Search Results

Search found 2045 results on 82 pages for 'james century'.

Page 12/82 | < Previous Page | 8 9 10 11 12 13 14 15 16 17 18 19  | Next Page >

  • More Fun with C# Iterators and Generators

    - by James Michael Hare
    In my last post, I talked quite a bit about iterators and how they can be really powerful tools for filtering a list of items down to a subset of items.  This had both pros and cons over returning a full collection, which, in summary, were:   Pros: If traversal is only partial, does not have to visit rest of collection. If evaluation method is costly, only incurs that cost on elements visited. Adds little to no garbage collection pressure.    Cons: Very slight performance impact if you know caller will always consume all items in collection. And as we saw in the last post, that con for the cost was very, very small and only really became evident on very tight loops consuming very large lists completely.    One of the key items to note, though, is the garbage!  In the traditional (return a new collection) method, if you have a 1,000,000 element collection, and wish to transform or filter it in some way, you have to allocate space for that copy of the collection.  That is, say you have a collection of 1,000,000 items and you want to double every item in the collection.  Well, that means you have to allocate a collection to hold those 1,000,000 items to return, which is a lot especially if you are just going to use it once and toss it.   Iterators, though, don't have this problem.  Each time you visit the node, it would return the doubled value of the node (in this example) and not allocate a second collection of 1,000,000 doubled items.  Do you see the distinction?  In both cases, we're consuming 1,000,000 items.  But in one case we pass back each doubled item which is just an int (for example's sake) on the stack and in the other case, we allocate a list containing 1,000,000 items which then must be garbage collected.   So iterators in C# are pretty cool, eh?  Well, here's one more thing a C# iterator can do that a traditional "return a new collection" transformation can't!   It can return **unbounded** collections!   I know, I know, that smells a lot like an infinite loop, eh?  Yes and no.  Basically, you're relying on the caller to put the bounds on the list, and as long as the caller doesn't you keep going.  Consider this example:   public static class Fibonacci {     // returns the infinite fibonacci sequence     public static IEnumerable<int> Sequence()     {         int iteration = 0;         int first = 1;         int second = 1;         int current = 0;         while (true)         {             if (iteration++ < 2)             {                 current = 1;             }             else             {                 current = first + second;                 second = first;                 first = current;             }             yield return current;         }     } }   Whoa, you say!  Yes, that's an infinite loop!  What the heck is going on there?  Yes, that was intentional.  Would it be better to have a fibonacci sequence that returns only a specific number of items?  Perhaps, but that wouldn't give you the power to defer the execution to the caller.   The beauty of this function is it is as infinite as the sequence itself!  The fibonacci sequence is unbounded, and so is this method.  It will continue to return fibonacci numbers for as long as you ask for them.  Now that's not something you can do with a traditional method that would return a collection of ints representing each number.  In that case you would eventually run out of memory as you got to higher and higher numbers.  This method, though, never runs out of memory.   Now, that said, you do have to know when you use it that it is an infinite collection and bound it appropriately.  Fortunately, Linq provides a lot of these extension methods for you!   Let's say you only want the first 10 fibonacci numbers:       foreach(var fib in Fibonacci.Sequence().Take(10))     {         Console.WriteLine(fib);     }   Or let's say you only want the fibonacci numbers that are less than 100:       foreach(var fib in Fibonacci.Sequence().TakeWhile(f => f < 100))     {         Console.WriteLine(fib);     }   So, you see, one of the nice things about iterators is their power to work with virtually any size (even infinite) collections without adding the garbage collection overhead of making new collections.    You can also do fun things like this to make a more "fluent" interface for for loops:   // A set of integer generator extension methods public static class IntExtensions {     // Begins counting to inifity, use To() to range this.     public static IEnumerable<int> Every(this int start)     {         // deliberately avoiding condition because keeps going         // to infinity for as long as values are pulled.         for (var i = start; ; ++i)         {             yield return i;         }     }     // Begins counting to infinity by the given step value, use To() to     public static IEnumerable<int> Every(this int start, int byEvery)     {         // deliberately avoiding condition because keeps going         // to infinity for as long as values are pulled.         for (var i = start; ; i += byEvery)         {             yield return i;         }     }     // Begins counting to inifity, use To() to range this.     public static IEnumerable<int> To(this int start, int end)     {         for (var i = start; i <= end; ++i)         {             yield return i;         }     }     // Ranges the count by specifying the upper range of the count.     public static IEnumerable<int> To(this IEnumerable<int> collection, int end)     {         return collection.TakeWhile(item => item <= end);     } }   Note that there are two versions of each method.  One that starts with an int and one that starts with an IEnumerable<int>.  This is to allow more power in chaining from either an existing collection or from an int.  This lets you do things like:   // count from 1 to 30 foreach(var i in 1.To(30)) {     Console.WriteLine(i); }     // count from 1 to 10 by 2s foreach(var i in 0.Every(2).To(10)) {     Console.WriteLine(i); }     // or, if you want an infinite sequence counting by 5s until something inside breaks you out... foreach(var i in 0.Every(5)) {     if (someCondition)     {         break;     }     ... }     Yes, those are kinda play functions and not particularly useful, but they show some of the power of generators and extension methods to form a fluid interface.   So what do you think?  What are some of your favorite generators and iterators?

    Read the article

  • ANTS Memory Profiler 7.0

    - by James Michael Hare
    I had always been a fan of ANTS products (Reflector is absolutely invaluable, and their performance profiler is great as well – very easy to use!), so I was curious to see what the ANTS Memory Profiler could show me. Background While a performance profiler will track how much time is typically spent in each unit of code, a memory profiler gives you much more detail on how and where your memory is being consumed and released in a program. As an example, I’d been working on a data access layer at work to call a market data web service.  This web service would take a list of symbols to quote and would return back the quote data.  To help consolidate the thousands of web requests per second we get and reduce load on the web services, we implemented a 5-second cache of quote data.  Not quite long enough to where customers will typically notice a quote go “stale”, but just long enough to be able to collapse multiple quote requests for the same symbol in a short period of time. A 5-second cache may not sound like much, but it actually pays off by saving us roughly 42% of our web service calls, while still providing relatively up-to-date information.  The question is whether or not the extra memory involved in maintaining the cache was worth it, so I decided to fire up the ANTS Memory Profiler and take a look at memory usage. First Impressions The main thing I’ve always loved about the ANTS tools is their ease of use.  Pretty much everything is right there in front of you in a way that makes it easy for you to find what you need with little digging required.  I’ve worked with other, older profilers before (that shall remain nameless other than to hint it was created by a very large chip maker) where it was a mind boggling experience to figure out how to do simple tasks. Not so with AMP.  The opening dialog is very straightforward.  You can choose from here whether to debug an executable, a web application (either in IIS or from VS’s web development server), windows services, etc. So I chose a .NET Executable and navigated to the build location of my test harness.  Then began profiling. At this point while the application is running, you can see a chart of the memory as it ebbs and wanes with allocations and collections.  At any given point in time, you can take snapshots (to compare states) zoom in, or choose to stop at any time.  Snapshots Taking a snapshot also gives you a breakdown of the managed memory heaps for each generation so you get an idea how many objects are staying around for extended periods of time (as an object lives and survives collections, it gets promoted into higher generations where collection becomes less frequent). Generating a snapshot brings up an analysis view with very handy graphs that show your generation sizes.  Almost all my memory is in Generation 1 in the managed memory component of the first graph, which is good news to me, because Gen 2 collections are much rarer.  I once3 made the mistake once of caching data for 30 minutes and found it didn’t get collected very quick after I released my reference because it had been promoted to Gen 2 – doh! Analysis It looks like (from the second pie chart) that the majority of the allocations were in the string class.  This also is expected for me because the majority of the memory allocated is in the web service responses, so it doesn’t seem the entities I’m adapting to (to prevent being too tightly coupled to the web service proxy classes, which can change easily out from under me) aren’t taking a significant portion of memory. I also appreciate that they have clear summary text in key places such as “No issues with large object heap fragmentation were detected”.  For novice users, this type of summary information can be critical to getting them to use a tool and develop a good working knowledge of it. There is also a handy link at the bottom for “What to look for on the summary” which loads a web page of help on key points to look for. Clicking over to the session overview, it’s easy to compare the samples at each snapshot to see how your memory is growing, shrinking, or staying relatively the same.  Looking at my snapshots, I’m pretty happy with the fact that memory allocation and heap size seems to be fairly stable and in control: Once again, you can check on the large object heap, generation one heap, and generation two heap across each snapshot to spot trends. Back on the analysis tab, we can go to the [Class List] button to get an idea what classes are making up the majority of our memory usage.  As was little surprise to me, System.String was the clear majority of my allocations, though I found it surprising that the System.Reflection.RuntimeMehtodInfo came in second.  I was curious about this, so I selected it and went into the [Instance Categorizer].  This view let me see where these instances to RuntimeMehtodInfo were coming from. So I scrolled back through the graph, and discovered that these were being held by the System.ServiceModel.ChannelFactoryRefCache and I was satisfied this was just an artifact of my WCF proxy. I also like that down at the bottom of the Instance Categorizer it gives you a series of filters and offers to guide you on which filter to use based on the problem you are trying to find.  For example, if I suspected a memory leak, I might try to filter for survivors in growing classes.  This means that for instances of a class that are growing in memory (more are being created than cleaned up), which ones are survivors (not collected) from garbage collection.  This might allow me to drill down and find places where I’m holding onto references by mistake and not freeing them! Finally, if you want to really see all your instances and who is holding onto them (preventing collection), you can go to the “Instance Retention Graph” which creates a graph showing what references are being held in memory and who is holding onto them. Visual Studio Integration Of course, VS has its own profiler built in – and for a free bundled profiler it is quite capable – but AMP gives a much cleaner and easier-to-use experience, and when you install it you also get the option of letting it integrate directly into VS. So once you go back into VS after installation, you’ll notice an ANTS menu which lets you launch the ANTS profiler directly from Visual Studio.   Clicking on one of these options fires up the project in the profiler immediately, allowing you to get right in.  It doesn’t integrate with the Visual Studio windows themselves (like the VS profiler does), but still the plethora of information it provides and the clear and concise manner in which it presents it makes it well worth it. Summary If you like the ANTS series of tools, you shouldn’t be disappointed with the ANTS Memory Profiler.  It was so easy to use that I was able to jump in with very little product knowledge and get the information I was looking it for. I’ve used other profilers before that came with 3-inch thick tomes that you had to read in order to get anywhere with the tool, and this one is not like that at all.  It’s built for your everyday developer to get in and find their problems quickly, and I like that! Tweet Technorati Tags: Influencers,ANTS,Memory,Profiler

    Read the article

  • Survey: Do you write custom SQL CLR procedures/functions/etc

    - by James Luetkehoelter
    I'm quite curious because despite the great capabilities of writing CLR-based stored procedures to off-load those nasty operations TSQL isn't that great at (like iteration, or complex math), I'm continuing to see a wealth of SQL 2008 databases with complex stored procedures and functions which would make great candidates. The in-house skill to create the CLR code exists as well, but there is flat out resistance to use it. In one scenario I was told "Oh, iteration isn't a problem because we've trained...(read more)

    Read the article

  • Debugging/Logging Techniques for End Users

    - by James Burgess
    I searched a bit, but didn't find anything particularly pertinent to my problem - so please do excuse me if I missed something! A few months back I inherited the source to a fairly-popular indie game project and have been working, along with another developer, on the code-base. We recently made our first release since taking over the development but we're a little stuck. A few users are experiencing slowdowns/lagging in the current version, as compared to the previous version, and we are not able to reproduce these issues in any of our various development environments (debug, release, different OSes, different machines, etc.). What I'd like to know is how can we go about implementing some form of logging/debugging mechanism into the game, that users can enable and send the reports to us for examination? We're not able to distribute debug binaries using the MSVS 2010 runtimes, due to the licensing - and wouldn't want to, for a variety of reasons. We'd really like to get to the bottom of this issue, even if just to find out it's nothing to do with our code base but everything to do with their system configuration. At the moment, we just have no leads - and the community isn't a very technically-savvy one, so we're unable to rely on 'expert' bug reports or investigations. I've seen the debug logging mechanism used in other applications and games for everything from logging simple errors to crash dumps. We're really at a loss at this stage as to how to address these issues, having been over every commit to the repository from the previous to the current version and not finding any real issues.

    Read the article

  • Easy user management on html site?

    - by James Buldon
    I hope I'm not asking a question for which the answer is obvious...If I am, apologies. Within my html site (i.e. not Wordpress, Joomla, etc.) I want to be able to have a level of user management. That means that some pages I want to be only accessible to certain people with the correct username and password. What's the best way to do this? Are there any available scripts out there? I guess I'm looking for a free/open source version of something like this: http://www.webassist.com/php-scripts-and-solutions/user-registration/

    Read the article

  • C#/.NET Little Wonders: The Generic Func Delegates

    - by James Michael Hare
    Once again, in this series of posts I look at the parts of the .NET Framework that may seem trivial, but can help improve your code by making it easier to write and maintain. The index of all my past little wonders posts can be found here. Back in one of my three original “Little Wonders” Trilogy of posts, I had listed generic delegates as one of the Little Wonders of .NET.  Later, someone posted a comment saying said that they would love more detail on the generic delegates and their uses, since my original entry just scratched the surface of them. Last week, I began our look at some of the handy generic delegates built into .NET with a description of delegates in general, and the Action family of delegates.  For this week, I’ll launch into a look at the Func family of generic delegates and how they can be used to support generic, reusable algorithms and classes. Quick Delegate Recap Delegates are similar to function pointers in C++ in that they allow you to store a reference to a method.  They can store references to either static or instance methods, and can actually be used to chain several methods together in one delegate. Delegates are very type-safe and can be satisfied with any standard method, anonymous method, or a lambda expression.  They can also be null as well (refers to no method), so care should be taken to make sure that the delegate is not null before you invoke it. Delegates are defined using the keyword delegate, where the delegate’s type name is placed where you would typically place the method name: 1: // This delegate matches any method that takes string, returns nothing 2: public delegate void Log(string message); This delegate defines a delegate type named Log that can be used to store references to any method(s) that satisfies its signature (whether instance, static, lambda expression, etc.). Delegate instances then can be assigned zero (null) or more methods using the operator = which replaces the existing delegate chain, or by using the operator += which adds a method to the end of a delegate chain: 1: // creates a delegate instance named currentLogger defaulted to Console.WriteLine (static method) 2: Log currentLogger = Console.Out.WriteLine; 3:  4: // invokes the delegate, which writes to the console out 5: currentLogger("Hi Standard Out!"); 6:  7: // append a delegate to Console.Error.WriteLine to go to std error 8: currentLogger += Console.Error.WriteLine; 9:  10: // invokes the delegate chain and writes message to std out and std err 11: currentLogger("Hi Standard Out and Error!"); While delegates give us a lot of power, it can be cumbersome to re-create fairly standard delegate definitions repeatedly, for this purpose the generic delegates were introduced in various stages in .NET.  These support various method types with particular signatures. Note: a caveat with generic delegates is that while they can support multiple parameters, they do not match methods that contains ref or out parameters. If you want to a delegate to represent methods that takes ref or out parameters, you will need to create a custom delegate. We’ve got the Func… delegates Just like it’s cousin, the Action delegate family, the Func delegate family gives us a lot of power to use generic delegates to make classes and algorithms more generic.  Using them keeps us from having to define a new delegate type when need to make a class or algorithm generic. Remember that the point of the Action delegate family was to be able to perform an “action” on an item, with no return results.  Thus Action delegates can be used to represent most methods that take 0 to 16 arguments but return void.  You can assign a method The Func delegate family was introduced in .NET 3.5 with the advent of LINQ, and gives us the power to define a function that can be called on 0 to 16 arguments and returns a result.  Thus, the main difference between Action and Func, from a delegate perspective, is that Actions return nothing, but Funcs return a result. The Func family of delegates have signatures as follows: Func<TResult> – matches a method that takes no arguments, and returns value of type TResult. Func<T, TResult> – matches a method that takes an argument of type T, and returns value of type TResult. Func<T1, T2, TResult> – matches a method that takes arguments of type T1 and T2, and returns value of type TResult. Func<T1, T2, …, TResult> – and so on up to 16 arguments, and returns value of type TResult. These are handy because they quickly allow you to be able to specify that a method or class you design will perform a function to produce a result as long as the method you specify meets the signature. For example, let’s say you were designing a generic aggregator, and you wanted to allow the user to define how the values will be aggregated into the result (i.e. Sum, Min, Max, etc…).  To do this, we would ask the user of our class to pass in a method that would take the current total, the next value, and produce a new total.  A class like this could look like: 1: public sealed class Aggregator<TValue, TResult> 2: { 3: // holds method that takes previous result, combines with next value, creates new result 4: private Func<TResult, TValue, TResult> _aggregationMethod; 5:  6: // gets or sets the current result of aggregation 7: public TResult Result { get; private set; } 8:  9: // construct the aggregator given the method to use to aggregate values 10: public Aggregator(Func<TResult, TValue, TResult> aggregationMethod = null) 11: { 12: if (aggregationMethod == null) throw new ArgumentNullException("aggregationMethod"); 13:  14: _aggregationMethod = aggregationMethod; 15: } 16:  17: // method to add next value 18: public void Aggregate(TValue nextValue) 19: { 20: // performs the aggregation method function on the current result and next and sets to current result 21: Result = _aggregationMethod(Result, nextValue); 22: } 23: } Of course, LINQ already has an Aggregate extension method, but that works on a sequence of IEnumerable<T>, whereas this is designed to work more with aggregating single results over time (such as keeping track of a max response time for a service). We could then use this generic aggregator to find the sum of a series of values over time, or the max of a series of values over time (among other things): 1: // creates an aggregator that adds the next to the total to sum the values 2: var sumAggregator = new Aggregator<int, int>((total, next) => total + next); 3:  4: // creates an aggregator (using static method) that returns the max of previous result and next 5: var maxAggregator = new Aggregator<int, int>(Math.Max); So, if we were timing the response time of a web method every time it was called, we could pass that response time to both of these aggregators to get an idea of the total time spent in that web method, and the max time spent in any one call to the web method: 1: // total will be 13 and max 13 2: int responseTime = 13; 3: sumAggregator.Aggregate(responseTime); 4: maxAggregator.Aggregate(responseTime); 5:  6: // total will be 20 and max still 13 7: responseTime = 7; 8: sumAggregator.Aggregate(responseTime); 9: maxAggregator.Aggregate(responseTime); 10:  11: // total will be 40 and max now 20 12: responseTime = 20; 13: sumAggregator.Aggregate(responseTime); 14: maxAggregator.Aggregate(responseTime); The Func delegate family is useful for making generic algorithms and classes, and in particular allows the caller of the method or user of the class to specify a function to be performed in order to generate a result. What is the result of a Func delegate chain? If you remember, we said earlier that you can assign multiple methods to a delegate by using the += operator to chain them.  So how does this affect delegates such as Func that return a value, when applied to something like the code below? 1: Func<int, int, int> combo = null; 2:  3: // What if we wanted to aggregate the sum and max together? 4: combo += (total, next) => total + next; 5: combo += Math.Max; 6:  7: // what is the result? 8: var comboAggregator = new Aggregator<int, int>(combo); Well, in .NET if you chain multiple methods in a delegate, they will all get invoked, but the result of the delegate is the result of the last method invoked in the chain.  Thus, this aggregator would always result in the Math.Max() result.  The other chained method (the sum) gets executed first, but it’s result is thrown away: 1: // result is 13 2: int responseTime = 13; 3: comboAggregator.Aggregate(responseTime); 4:  5: // result is still 13 6: responseTime = 7; 7: comboAggregator.Aggregate(responseTime); 8:  9: // result is now 20 10: responseTime = 20; 11: comboAggregator.Aggregate(responseTime); So remember, you can chain multiple Func (or other delegates that return values) together, but if you do so you will only get the last executed result. Func delegates and co-variance/contra-variance in .NET 4.0 Just like the Action delegate, as of .NET 4.0, the Func delegate family is contra-variant on its arguments.  In addition, it is co-variant on its return type.  To support this, in .NET 4.0 the signatures of the Func delegates changed to: Func<out TResult> – matches a method that takes no arguments, and returns value of type TResult (or a more derived type). Func<in T, out TResult> – matches a method that takes an argument of type T (or a less derived type), and returns value of type TResult(or a more derived type). Func<in T1, in T2, out TResult> – matches a method that takes arguments of type T1 and T2 (or less derived types), and returns value of type TResult (or a more derived type). Func<in T1, in T2, …, out TResult> – and so on up to 16 arguments, and returns value of type TResult (or a more derived type). Notice the addition of the in and out keywords before each of the generic type placeholders.  As we saw last week, the in keyword is used to specify that a generic type can be contra-variant -- it can match the given type or a type that is less derived.  However, the out keyword, is used to specify that a generic type can be co-variant -- it can match the given type or a type that is more derived. On contra-variance, if you are saying you need an function that will accept a string, you can just as easily give it an function that accepts an object.  In other words, if you say “give me an function that will process dogs”, I could pass you a method that will process any animal, because all dogs are animals.  On the co-variance side, if you are saying you need a function that returns an object, you can just as easily pass it a function that returns a string because any string returned from the given method can be accepted by a delegate expecting an object result, since string is more derived.  Once again, in other words, if you say “give me a method that creates an animal”, I can pass you a method that will create a dog, because all dogs are animals. It really all makes sense, you can pass a more specific thing to a less specific parameter, and you can return a more specific thing as a less specific result.  In other words, pay attention to the direction the item travels (parameters go in, results come out).  Keeping that in mind, you can always pass more specific things in and return more specific things out. For example, in the code below, we have a method that takes a Func<object> to generate an object, but we can pass it a Func<string> because the return type of object can obviously accept a return value of string as well: 1: // since Func<object> is co-variant, this will access Func<string>, etc... 2: public static string Sequence(int count, Func<object> generator) 3: { 4: var builder = new StringBuilder(); 5:  6: for (int i=0; i<count; i++) 7: { 8: object value = generator(); 9: builder.Append(value); 10: } 11:  12: return builder.ToString(); 13: } Even though the method above takes a Func<object>, we can pass a Func<string> because the TResult type placeholder is co-variant and accepts types that are more derived as well: 1: // delegate that's typed to return string. 2: Func<string> stringGenerator = () => DateTime.Now.ToString(); 3:  4: // This will work in .NET 4.0, but not in previous versions 5: Sequence(100, stringGenerator); Previous versions of .NET implemented some forms of co-variance and contra-variance before, but .NET 4.0 goes one step further and allows you to pass or assign an Func<A, BResult> to a Func<Y, ZResult> as long as A is less derived (or same) as Y, and BResult is more derived (or same) as ZResult. Sidebar: The Func and the Predicate A method that takes one argument and returns a bool is generally thought of as a predicate.  Predicates are used to examine an item and determine whether that item satisfies a particular condition.  Predicates are typically unary, but you may also have binary and other predicates as well. Predicates are often used to filter results, such as in the LINQ Where() extension method: 1: var numbers = new[] { 1, 2, 4, 13, 8, 10, 27 }; 2:  3: // call Where() using a predicate which determines if the number is even 4: var evens = numbers.Where(num => num % 2 == 0); As of .NET 3.5, predicates are typically represented as Func<T, bool> where T is the type of the item to examine.  Previous to .NET 3.5, there was a Predicate<T> type that tended to be used (which we’ll discuss next week) and is still supported, but most developers recommend using Func<T, bool> now, as it prevents confusion with overloads that accept unary predicates and binary predicates, etc.: 1: // this seems more confusing as an overload set, because of Predicate vs Func 2: public static SomeMethod(Predicate<int> unaryPredicate) { } 3: public static SomeMethod(Func<int, int, bool> binaryPredicate) { } 4:  5: // this seems more consistent as an overload set, since just uses Func 6: public static SomeMethod(Func<int, bool> unaryPredicate) { } 7: public static SomeMethod(Func<int, int, bool> binaryPredicate) { } Also, even though Predicate<T> and Func<T, bool> match the same signatures, they are separate types!  Thus you cannot assign a Predicate<T> instance to a Func<T, bool> instance and vice versa: 1: // the same method, lambda expression, etc can be assigned to both 2: Predicate<int> isEven = i => (i % 2) == 0; 3: Func<int, bool> alsoIsEven = i => (i % 2) == 0; 4:  5: // but the delegate instances cannot be directly assigned, strongly typed! 6: // ERROR: cannot convert type... 7: isEven = alsoIsEven; 8:  9: // however, you can assign by wrapping in a new instance: 10: isEven = new Predicate<int>(alsoIsEven); 11: alsoIsEven = new Func<int, bool>(isEven); So, the general advice that seems to come from most developers is that Predicate<T> is still supported, but we should use Func<T, bool> for consistency in .NET 3.5 and above. Sidebar: Func as a Generator for Unit Testing One area of difficulty in unit testing can be unit testing code that is based on time of day.  We’d still want to unit test our code to make sure the logic is accurate, but we don’t want the results of our unit tests to be dependent on the time they are run. One way (of many) around this is to create an internal generator that will produce the “current” time of day.  This would default to returning result from DateTime.Now (or some other method), but we could inject specific times for our unit testing.  Generators are typically methods that return (generate) a value for use in a class/method. For example, say we are creating a CacheItem<T> class that represents an item in the cache, and we want to make sure the item shows as expired if the age is more than 30 seconds.  Such a class could look like: 1: // responsible for maintaining an item of type T in the cache 2: public sealed class CacheItem<T> 3: { 4: // helper method that returns the current time 5: private static Func<DateTime> _timeGenerator = () => DateTime.Now; 6:  7: // allows internal access to the time generator 8: internal static Func<DateTime> TimeGenerator 9: { 10: get { return _timeGenerator; } 11: set { _timeGenerator = value; } 12: } 13:  14: // time the item was cached 15: public DateTime CachedTime { get; private set; } 16:  17: // the item cached 18: public T Value { get; private set; } 19:  20: // item is expired if older than 30 seconds 21: public bool IsExpired 22: { 23: get { return _timeGenerator() - CachedTime > TimeSpan.FromSeconds(30.0); } 24: } 25:  26: // creates the new cached item, setting cached time to "current" time 27: public CacheItem(T value) 28: { 29: Value = value; 30: CachedTime = _timeGenerator(); 31: } 32: } Then, we can use this construct to unit test our CacheItem<T> without any time dependencies: 1: var baseTime = DateTime.Now; 2:  3: // start with current time stored above (so doesn't drift) 4: CacheItem<int>.TimeGenerator = () => baseTime; 5:  6: var target = new CacheItem<int>(13); 7:  8: // now add 15 seconds, should still be non-expired 9: CacheItem<int>.TimeGenerator = () => baseTime.AddSeconds(15); 10:  11: Assert.IsFalse(target.IsExpired); 12:  13: // now add 31 seconds, should now be expired 14: CacheItem<int>.TimeGenerator = () => baseTime.AddSeconds(31); 15:  16: Assert.IsTrue(target.IsExpired); Now we can unit test for 1 second before, 1 second after, 1 millisecond before, 1 day after, etc.  Func delegates can be a handy tool for this type of value generation to support more testable code.  Summary Generic delegates give us a lot of power to make truly generic algorithms and classes.  The Func family of delegates is a great way to be able to specify functions to calculate a result based on 0-16 arguments.  Stay tuned in the weeks that follow for other generic delegates in the .NET Framework!   Tweet Technorati Tags: .NET, C#, CSharp, Little Wonders, Generics, Func, Delegates

    Read the article

  • Amnesia doesn't start due to audio problems

    - by james
    I have a problem with amnesia game. After Intro and clicking continue button few times, when game is supposed to start it crashes. Here is console output: ALSA lib pcm_dmix.c:1018:(snd_pcm_dmix_open) unable to open slave ALSA lib pcm.c:2217:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.rear ALSA lib pcm.c:2217:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.center_lfe ALSA lib pcm.c:2217:(snd_pcm_open_noupdate) Unknown PCM cards.pcm.side ALSA lib audio/pcm_bluetooth.c:1614:(audioservice_expect) BT_GET_CAPABILITIES failed : Input/output error(5) ALSA lib audio/pcm_bluetooth.c:1614:(audioservice_expect) BT_GET_CAPABILITIES failed : Input/output error(5) ALSA lib audio/pcm_bluetooth.c:1614:(audioservice_expect) BT_GET_CAPABILITIES failed : Input/output error(5) ALSA lib audio/pcm_bluetooth.c:1614:(audioservice_expect) BT_GET_CAPABILITIES failed : Input/output error(5) ALSA lib pcm_dmix.c:957:(snd_pcm_dmix_open) The dmix plugin supports only playback stream ALSA lib pcm_dmix.c:1018:(snd_pcm_dmix_open) unable to open slave Cannot connect to server socket err = No such file or directory Cannot connect to server socket jack server is not running or cannot be started I should mention I have integrated both graphic and sound card.

    Read the article

  • Is there any reason to allow Yahoo! Slurp to crawl my site?

    - by James Skemp
    I thought a year or more ago Yahoo! would be using another search engine for results, and no longer using their own Slurp bot. However, a couple of the sites I manage Yahoo! Slurp continues to crawl pages, and seems to ignore the Gone status code when returned (as it keeps coming back). Is there any reason why I wouldn't want to block Yahoo! Slurp via robots.txt or by IP (since it tends to ignore robots.txt in some cases anyways)? I've confirmed that when the bot does hit it is from Yahoo! IPs, so I believe this is a legit instance of the bot. Is Yahoo Search the same as Bing Search now? is a related question, but I don't think it completely answers whether one should add a new block of the bot.

    Read the article

  • C#/.NET Little Wonders: The Nullable static class

    - by James Michael Hare
    Once again, in this series of posts I look at the parts of the .NET Framework that may seem trivial, but can help improve your code by making it easier to write and maintain. The index of all my past little wonders posts can be found here. Today we’re going to look at an interesting Little Wonder that can be used to mitigate what could be considered a Little Pitfall.  The Little Wonder we’ll be examining is the System.Nullable static class.  No, not the System.Nullable<T> class, but a static helper class that has one useful method in particular that we will examine… but first, let’s look at the Little Pitfall that makes this wonder so useful. Little Pitfall: Comparing nullable value types using <, >, <=, >= Examine this piece of code, without examining it too deeply, what’s your gut reaction as to the result? 1: int? x = null; 2:  3: if (x < 100) 4: { 5: Console.WriteLine("True, {0} is less than 100.", 6: x.HasValue ? x.ToString() : "null"); 7: } 8: else 9: { 10: Console.WriteLine("False, {0} is NOT less than 100.", 11: x.HasValue ? x.ToString() : "null"); 12: } Your gut would be to say true right?  It would seem to make sense that a null integer is less than the integer constant 100.  But the result is actually false!  The null value is not less than 100 according to the less-than operator. It looks even more outrageous when you consider this also evaluates to false: 1: int? x = null; 2:  3: if (x < int.MaxValue) 4: { 5: // ... 6: } So, are we saying that null is less than every valid int value?  If that were true, null should be less than int.MinValue, right?  Well… no: 1: int? x = null; 2:  3: // um... hold on here, x is NOT less than min value? 4: if (x < int.MinValue) 5: { 6: // ... 7: } So what’s going on here?  If we use greater than instead of less than, we see the same little dilemma: 1: int? x = null; 2:  3: // once again, null is not greater than anything either... 4: if (x > int.MinValue) 5: { 6: // ... 7: } It turns out that four of the comparison operators (<, <=, >, >=) are designed to return false anytime at least one of the arguments is null when comparing System.Nullable wrapped types that expose the comparison operators (short, int, float, double, DateTime, TimeSpan, etc.).  What’s even odder is that even though the two equality operators (== and !=) work correctly, >= and <= have the same issue as < and > and return false if both System.Nullable wrapped operator comparable types are null! 1: DateTime? x = null; 2: DateTime? y = null; 3:  4: if (x <= y) 5: { 6: Console.WriteLine("You'd think this is true, since both are null, but it's not."); 7: } 8: else 9: { 10: Console.WriteLine("It's false because <=, <, >, >= don't work on null."); 11: } To make matters even more confusing, take for example your usual check to see if something is less than, greater to, or equal: 1: int? x = null; 2: int? y = 100; 3:  4: if (x < y) 5: { 6: Console.WriteLine("X is less than Y"); 7: } 8: else if (x > y) 9: { 10: Console.WriteLine("X is greater than Y"); 11: } 12: else 13: { 14: // We fall into the "equals" assumption, but clearly null != 100! 15: Console.WriteLine("X is equal to Y"); 16: } Yes, this code outputs “X is equal to Y” because both the less-than and greater-than operators return false when a Nullable wrapped operator comparable type is null.  This violates a lot of our assumptions because we assume is something is not less than something, and it’s not greater than something, it must be equal.  So keep in mind, that the only two comparison operators that work on Nullable wrapped types where at least one is null are the equals (==) and not equals (!=) operators: 1: int? x = null; 2: int? y = 100; 3:  4: if (x == y) 5: { 6: Console.WriteLine("False, x is null, y is not."); 7: } 8:  9: if (x != y) 10: { 11: Console.WriteLine("True, x is null, y is not."); 12: } Solution: The Nullable static class So we’ve seen that <, <=, >, and >= have some interesting and perhaps unexpected behaviors that can trip up a novice developer who isn’t expecting the kinks that System.Nullable<T> types with comparison operators can throw.  How can we easily mitigate this? Well, obviously, you could do null checks before each check, but that starts to get ugly: 1: if (x.HasValue) 2: { 3: if (y.HasValue) 4: { 5: if (x < y) 6: { 7: Console.WriteLine("x < y"); 8: } 9: else if (x > y) 10: { 11: Console.WriteLine("x > y"); 12: } 13: else 14: { 15: Console.WriteLine("x == y"); 16: } 17: } 18: else 19: { 20: Console.WriteLine("x > y because y is null and x isn't"); 21: } 22: } 23: else if (y.HasValue) 24: { 25: Console.WriteLine("x < y because x is null and y isn't"); 26: } 27: else 28: { 29: Console.WriteLine("x == y because both are null"); 30: } Yes, we could probably simplify this logic a bit, but it’s still horrendous!  So what do we do if we want to consider null less than everything and be able to properly compare Nullable<T> wrapped value types? The key is the System.Nullable static class.  This class is a companion class to the System.Nullable<T> class and allows you to use a few helper methods for Nullable<T> wrapped types, including a static Compare<T>() method of the. What’s so big about the static Compare<T>() method?  It implements an IComparer compatible comparison on Nullable<T> types.  Why do we care?  Well, if you look at the MSDN description for how IComparer works, you’ll read: Comparing null with any type is allowed and does not generate an exception when using IComparable. When sorting, null is considered to be less than any other object. This is what we probably want!  We want null to be less than everything!  So now we can change our logic to use the Nullable.Compare<T>() static method: 1: int? x = null; 2: int? y = 100; 3:  4: if (Nullable.Compare(x, y) < 0) 5: { 6: // Yes! x is null, y is not, so x is less than y according to Compare(). 7: Console.WriteLine("x < y"); 8: } 9: else if (Nullable.Compare(x, y) > 0) 10: { 11: Console.WriteLine("x > y"); 12: } 13: else 14: { 15: Console.WriteLine("x == y"); 16: } Summary So, when doing math comparisons between two numeric values where one of them may be a null Nullable<T>, consider using the System.Nullable.Compare<T>() method instead of the comparison operators.  It will treat null less than any value, and will avoid logic consistency problems when relying on < returning false to indicate >= is true and so on. Tweet   Technorati Tags: C#,C-Sharp,.NET,Little Wonders,Little Pitfalls,Nulalble

    Read the article

  • C#/.NET Little Wonders: Use Cast() and TypeOf() to Change Sequence Type

    - by James Michael Hare
    Once again, in this series of posts I look at the parts of the .NET Framework that may seem trivial, but can help improve your code by making it easier to write and maintain. The index of all my past little wonders posts can be found here. We’ve seen how the Select() extension method lets you project a sequence from one type to a new type which is handy for getting just parts of items, or building new items.  But what happens when the items in the sequence are already the type you want, but the sequence itself is typed to an interface or super-type instead of the sub-type you need? For example, you may have a sequence of Rectangle stored in an IEnumerable<Shape> and want to consider it an IEnumerable<Rectangle> sequence instead.  Today we’ll look at two handy extension methods, Cast<TResult>() and OfType<TResult>() which help you with this task. Cast<TResult>() – Attempt to cast all items to type TResult So, the first thing we can do would be to attempt to create a sequence of TResult from every item in the source sequence.  Typically we’d do this if we had an IEnumerable<T> where we knew that every item was actually a TResult where TResult inherits/implements T. For example, assume the typical Shape example classes: 1: // abstract base class 2: public abstract class Shape { } 3:  4: // a basic rectangle 5: public class Rectangle : Shape 6: { 7: public int Widtgh { get; set; } 8: public int Height { get; set; } 9: } And let’s assume we have a sequence of Shape where every Shape is a Rectangle… 1: var shapes = new List<Shape> 2: { 3: new Rectangle { Width = 3, Height = 5 }, 4: new Rectangle { Width = 10, Height = 13 }, 5: // ... 6: }; To get the sequence of Shape as a sequence of Rectangle, of course, we could use a Select() clause, such as: 1: // select each Shape, cast it to Rectangle 2: var rectangles = shapes 3: .Select(s => (Rectangle)s) 4: .ToList(); But that’s a bit verbose, and fortunately there is already a facility built in and ready to use in the form of the Cast<TResult>() extension method: 1: // cast each item to Rectangle and store in a List<Rectangle> 2: var rectangles = shapes 3: .Cast<Rectangle>() 4: .ToList(); However, we should note that if anything in the list cannot be cast to a Rectangle, you will get an InvalidCastException thrown at runtime.  Thus, if our Shape sequence had a Circle in it, the call to Cast<Rectangle>() would have failed.  As such, you should only do this when you are reasonably sure of what the sequence actually contains (or are willing to handle an exception if you’re wrong). Another handy use of Cast<TResult>() is using it to convert an IEnumerable to an IEnumerable<T>.  If you look at the signature, you’ll see that the Cast<TResult>() extension method actually extends the older, object-based IEnumerable interface instead of the newer, generic IEnumerable<T>.  This is your gateway method for being able to use LINQ on older, non-generic sequences.  For example, consider the following: 1: // the older, non-generic collections are sequence of object 2: var shapes = new ArrayList 3: { 4: new Rectangle { Width = 3, Height = 13 }, 5: new Rectangle { Width = 10, Height = 20 }, 6: // ... 7: }; Since this is an older, object based collection, we cannot use the LINQ extension methods on it directly.  For example, if I wanted to query the Shape sequence for only those Rectangles whose Width is > 5, I can’t do this: 1: // compiler error, Where() operates on IEnumerable<T>, not IEnumerable 2: var bigRectangles = shapes.Where(r => r.Width > 5); However, I can use Cast<Rectangle>() to treat my ArrayList as an IEnumerable<Rectangle> and then do the query! 1: // ah, that’s better! 2: var bigRectangles = shapes.Cast<Rectangle>().Where(r => r.Width > 5); Or, if you prefer, in LINQ query expression syntax: 1: var bigRectangles = from s in shapes.Cast<Rectangle>() 2: where s.Width > 5 3: select s; One quick warning: Cast<TResult>() only attempts to cast, it won’t perform a cast conversion.  That is, consider this: 1: var intList = new List<int> { 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89 }; 2:  3: // casting ints to longs, this should work, right? 4: var asLong = intList.Cast<long>().ToList(); Will the code above work?  No, you’ll get a InvalidCastException. Remember that Cast<TResult>() is an extension of IEnumerable, thus it is a sequence of object, which means that it will box every int as an object as it enumerates over it, and there is no cast conversion from object to long, and thus the cast fails.  In other words, a cast from int to long will succeed because there is a conversion from int to long.  But a cast from int to object to long will not, because you can only unbox an item by casting it to its exact type. For more information on why cast-converting boxed values doesn’t work, see this post on The Dangers of Casting Boxed Values (here). OfType<TResult>() – Filter sequence to only items of type TResult So, we’ve seen how we can use Cast<TResult>() to change the type of our sequence, when we expect all the items of the sequence to be of a specific type.  But what do we do when a sequence contains many different types, and we are only concerned with a subset of a given type? For example, what if a sequence of Shape contains Rectangle and Circle instances, and we just want to select all of the Rectangle instances?  Well, let’s say we had this sequence of Shape: 1: var shapes = new List<Shape> 2: { 3: new Rectangle { Width = 3, Height = 5 }, 4: new Rectangle { Width = 10, Height = 13 }, 5: new Circle { Radius = 10 }, 6: new Square { Side = 13 }, 7: // ... 8: }; Well, we could get the rectangles using Select(), like: 1: var onlyRectangles = shapes.Where(s => s is Rectangle).ToList(); But fortunately, an easier way has already been written for us in the form of the OfType<T>() extension method: 1: // returns only a sequence of the shapes that are Rectangles 2: var onlyRectangles = shapes.OfType<Rectangle>().ToList(); Now we have a sequence of only the Rectangles in the original sequence, we can also use this to chain other queries that depend on Rectangles, such as: 1: // select only Rectangles, then filter to only those more than 2: // 5 units wide... 3: var onlyBigRectangles = shapes.OfType<Rectangle>() 4: .Where(r => r.Width > 5) 5: .ToList(); The OfType<Rectangle>() will filter the sequence to only the items that are of type Rectangle (or a subclass of it), and that results in an IEnumerable<Rectangle>, we can then apply the other LINQ extension methods to query that list further. Just as Cast<TResult>() is an extension method on IEnumerable (and not IEnumerable<T>), the same is true for OfType<T>().  This means that you can use OfType<TResult>() on object-based collections as well. For example, given an ArrayList containing Shapes, as below: 1: // object-based collections are a sequence of object 2: var shapes = new ArrayList 3: { 4: new Rectangle { Width = 3, Height = 5 }, 5: new Rectangle { Width = 10, Height = 13 }, 6: new Circle { Radius = 10 }, 7: new Square { Side = 13 }, 8: // ... 9: }; We can use OfType<Rectangle> to filter the sequence to only Rectangle items (and subclasses), and then chain other LINQ expressions, since we will then be of type IEnumerable<Rectangle>: 1: // OfType() converts the sequence of object to a new sequence 2: // containing only Rectangle or sub-types of Rectangle. 3: var onlyBigRectangles = shapes.OfType<Rectangle>() 4: .Where(r => r.Width > 5) 5: .ToList(); Summary So now we’ve seen two different ways to get a sequence of a superclass or interface down to a more specific sequence of a subclass or implementation.  The Cast<TResult>() method casts every item in the source sequence to type TResult, and the OfType<TResult>() method selects only those items in the source sequence that are of type TResult. You can use these to downcast sequences, or adapt older types and sequences that only implement IEnumerable (such as DataTable, ArrayList, etc.). Technorati Tags: C#,CSharp,.NET,LINQ,Little Wonders,TypeOf,Cast,IEnumerable<T>

    Read the article

  • What is a best practice tier structure of a Java EE 6/7 application?

    - by James Drinkard
    I was attempting to find a best practice for modeling the tiers in a Java EE application yesterday and couldn't come up with anything current. In the past, say java 1.4, it was four tiers: Presentation Tier Web Tier Business Logic Tier DAL (Data Access Layer ) which I always considered a tier and not a layer. After working with Web Services and SOA I thought to add in a services tier, but that may fall under 3. the business logic tier. I did searches for quite a while and reading articles. It seems like Domain Driven Design is becoming more popular, but I couldn't find a diagram on it's tier structure. Anyone have ideas or diagrams on what the proper tier structure is for newer Java EE applications or is it really the same, but more items are ranked under the four I've mentioned?

    Read the article

  • How *not* to handle a compensation step on failure in an SSIS package

    - by James Luetkehoelter
    Just stumbed across this where I'm working. Someone created a global error handler for a package that included this SQL step: DELETE FROM Table WHERE DateDiff(MI, ExportedDate, GetDate()) < 5 So if the package runs for longer than 5 minutes and fails, nothing gets cleaned up. Please people, don't do this... Share this post: email it! | bookmark it! | digg it! | reddit! | kick it! | live it!...(read more)

    Read the article

  • SEO Keyword Research Help

    - by James
    Hi Everyone, I'm new at SEO and keyword research. I am using Market Samurai as my research tool, and I was wondering if I could ask for your help to identify the best key word to target for my niche. I do plan on incorporating all of them into my site, but I wanted to start with one. If you could give me your input on these keywords, I would appreciate it. This is all new to me :) I'm too new to post pictures, but here are my keywords (Searches, SEO Traffic, and SEO Value / Day): Searches | SEO Traffic | PBR | SEO Value | Average PR/Backlinks of Current Top 10 1: 730 | 307 | 20% | 2311.33 | 1.9 / 7k-60k 2: 325 | 137 | 24% | 822.94 | 2.3 / 7k-60k 3: 398 | 167 | 82% | 589.79 | 1.6 / 7k-60k I'm wondering if the PBR (Phrase-to-broad) value of #1 is too low. It seems like the best value because the SEOV is crazy high. That is like $70k a month. #3 has the highest PBR, but also the lowest SEOV. #2 doesn't seem worth it because of the PR competetion. Might be a little too hard to get into the top page of Google. I'm wondering which keywords to target, and if I should be looking at any other metric to see if this is a profitable niche to jump into. Thanks.

    Read the article

  • C#: A "Dumbed-Down" C++?

    - by James Michael Hare
    I was spending a lovely day this last weekend watching my sons play outside in one of the better weekends we've had here in Saint Louis for quite some time, and whilst watching them and making sure no limbs were broken or eyes poked out with sticks and other various potential injuries, I was perusing (in the correct sense of the word) this month's MSDN magazine to get a sense of the latest VS2010 features in both IDE and in languages. When I got to the back pages, I saw a wonderful article by David S. Platt entitled, "In Praise of Dumbing Down"  (msdn.microsoft.com/en-us/magazine/ee336129.aspx).  The title captivated me and I read it and found myself agreeing with it completely especially as it related to my first post on divorcing C++ as my favorite language. Unfortunately, as Mr. Platt mentions, the term dumbing-down has negative connotations, but is really and truly a good thing.  You are, in essence, taking something that is extremely complex and reducing it to something that is much easier to use and far less error prone.  Adding safeties to power tools and anti-kick mechanisms to chainsaws are in some sense "dumbing them down" to the common user -- but that also makes them safer and more accessible for the common user.  This was exactly my point with C++ and C#.  I did not mean to infer that C++ was not a useful or good language, but that in a very high percentage of cases, is too complex and error prone for the job at hand. Choosing the correct programming language for a job is a lot like choosing any other tool for a task.  For example: if I want to dig a French drain in my lawn, I can attempt to use a huge tractor-like backhoe and the job would be done far quicker than if I would dig it by hand.  I can't deny that the backhoe has the raw power and speed to perform.  But you also cannot deny that my chances of injury or chances of severing utility lines or other resources climb at an exponential rate inverse to the amount of training I may have on that machinery. Is C++ a powerful tool?  Oh yes, and it's great for those tasks where speed and performance are paramount.  But for most of us, it's the wrong tool.  And keep in mind, I say this even though I have 17 years of experience in using it and feel myself highly adept in utilizing its features both in the standard libraries, the STL, and in supplemental libraries such as BOOST.  Which, although greatly help with adding powerful features quickly, do very little to curb the relative dangers of the language. So, you may say, the fault is in the developer, that if the developer had some higher skills or if we only hired C++ experts this would not be an issue.  Now, I will concede there is some truth to this.  Obviously, the higher skilled C++ developers you hire the better the chance they will produce highly performant and error-free code.  However, what good is that to the average developer who cannot afford a full stable of C++ experts? That's my point with C#:  It's like a kinder, gentler C++.  It gives you nearly the same speed, and in many ways even more power than C++, and it gives you a much softer cushion for novices to fall against if they code less-than-optimally.  A bug is a bug, of course, in any language, but C# does a good job of hiding and taking on the task of handling almost all of the resource issues that make C++ so tricky.  For my money, C# is much more maintainable, more feature-rich, second only slightly in performance, faster to market, and -- last but not least -- safer and easier to use.  That's why, where I work, I much prefer to see the developers moving to C#.  The quantity of bugs is much lower, and we don't need to hire "experts" to achieve the same results since the language itself handles those resource pitfalls so prevalent in poorly written C++ code.  C++ will still have its place in the world, and I'm sure I'll still use it now and again where it is truly the correct tool for the job, but for nearly every other project C# is a wonderfully "dumbed-down" version of C++ -- in the very best sense -- and to me, that's the smart choice.

    Read the article

  • C#: Handling Notifications: inheritance, events, or delegates?

    - by James Michael Hare
    Often times as developers we have to design a class where we get notification when certain things happen. In older object-oriented code this would often be implemented by overriding methods -- with events, delegates, and interfaces, however, we have far more elegant options. So, when should you use each of these methods and what are their strengths and weaknesses? Now, for the purposes of this article when I say notification, I'm just talking about ways for a class to let a user know that something has occurred. This can be through any programmatic means such as inheritance, events, delegates, etc. So let's build some context. I'm sitting here thinking about a provider neutral messaging layer for the place I work, and I got to the point where I needed to design the message subscriber which will receive messages from the message bus. Basically, what we want is to be able to create a message listener and have it be called whenever a new message arrives. Now, back before the flood we would have done this via inheritance and an abstract class: 1:  2: // using inheritance - omitting argument null checks and halt logic 3: public abstract class MessageListener 4: { 5: private ISubscriber _subscriber; 6: private bool _isHalted = false; 7: private Thread _messageThread; 8:  9: // assign the subscriber and start the messaging loop 10: public MessageListener(ISubscriber subscriber) 11: { 12: _subscriber = subscriber; 13: _messageThread = new Thread(MessageLoop); 14: _messageThread.Start(); 15: } 16:  17: // user will override this to process their messages 18: protected abstract void OnMessageReceived(Message msg); 19:  20: // handle the looping in the thread 21: private void MessageLoop() 22: { 23: while(!_isHalted) 24: { 25: // as long as processing, wait 1 second for message 26: Message msg = _subscriber.Receive(TimeSpan.FromSeconds(1)); 27: if(msg != null) 28: { 29: OnMessageReceived(msg); 30: } 31: } 32: } 33: ... 34: } It seems so odd to write this kind of code now. Does it feel odd to you? Maybe it's just because I've gotten so used to delegation that I really don't like the feel of this. To me it is akin to saying that if I want to drive my car I need to derive a new instance of it just to put myself in the driver's seat. And yet, unquestionably, five years ago I would have probably written the code as you see above. To me, inheritance is a flawed approach for notifications due to several reasons: Inheritance is one of the HIGHEST forms of coupling. You can't seal the listener class because it depends on sub-classing to work. Because C# does not allow multiple-inheritance, I've spent my one inheritance implementing this class. Every time you need to listen to a bus, you have to derive a class which leads to lots of trivial sub-classes. The act of consuming a message should be a separate responsibility than the act of listening for a message (SRP). Inheritance is such a strong statement (this IS-A that) that it should only be used in building type hierarchies and not for overriding use-specific behaviors and notifications. Chances are, if a class needs to be inherited to be used, it most likely is not designed as well as it could be in today's modern programming languages. So lets look at the other tools available to us for getting notified instead. Here's a few other choices to consider. Have the listener expose a MessageReceived event. Have the listener accept a new IMessageHandler interface instance. Have the listener accept an Action<Message> delegate. Really, all of these are different forms of delegation. Now, .NET events are a bit heavier than the other types of delegates in terms of run-time execution, but they are a great way to allow others using your class to subscribe to your events: 1: // using event - ommiting argument null checks and halt logic 2: public sealed class MessageListener 3: { 4: private ISubscriber _subscriber; 5: private bool _isHalted = false; 6: private Thread _messageThread; 7:  8: // assign the subscriber and start the messaging loop 9: public MessageListener(ISubscriber subscriber) 10: { 11: _subscriber = subscriber; 12: _messageThread = new Thread(MessageLoop); 13: _messageThread.Start(); 14: } 15:  16: // user will override this to process their messages 17: public event Action<Message> MessageReceived; 18:  19: // handle the looping in the thread 20: private void MessageLoop() 21: { 22: while(!_isHalted) 23: { 24: // as long as processing, wait 1 second for message 25: Message msg = _subscriber.Receive(TimeSpan.FromSeconds(1)); 26: if(msg != null && MessageReceived != null) 27: { 28: MessageReceived(msg); 29: } 30: } 31: } 32: } Note, now we can seal the class to avoid changes and the user just needs to provide a message handling method: 1: theListener.MessageReceived += CustomReceiveMethod; However, personally I don't think events hold up as well in this case because events are largely optional. To me, what is the point of a listener if you create one with no event listeners? So in my mind, use events when handling the notification is optional. So how about the delegation via interface? I personally like this method quite a bit. Basically what it does is similar to inheritance method mentioned first, but better because it makes it easy to split the part of the class that doesn't change (the base listener behavior) from the part that does change (the user-specified action after receiving a message). So assuming we had an interface like: 1: public interface IMessageHandler 2: { 3: void OnMessageReceived(Message receivedMessage); 4: } Our listener would look like this: 1: // using delegation via interface - omitting argument null checks and halt logic 2: public sealed class MessageListener 3: { 4: private ISubscriber _subscriber; 5: private IMessageHandler _handler; 6: private bool _isHalted = false; 7: private Thread _messageThread; 8:  9: // assign the subscriber and start the messaging loop 10: public MessageListener(ISubscriber subscriber, IMessageHandler handler) 11: { 12: _subscriber = subscriber; 13: _handler = handler; 14: _messageThread = new Thread(MessageLoop); 15: _messageThread.Start(); 16: } 17:  18: // handle the looping in the thread 19: private void MessageLoop() 20: { 21: while(!_isHalted) 22: { 23: // as long as processing, wait 1 second for message 24: Message msg = _subscriber.Receive(TimeSpan.FromSeconds(1)); 25: if(msg != null) 26: { 27: _handler.OnMessageReceived(msg); 28: } 29: } 30: } 31: } And they would call it by creating a class that implements IMessageHandler and pass that instance into the constructor of the listener. I like that this alleviates the issues of inheritance and essentially forces you to provide a handler (as opposed to events) on construction. Well, this is good, but personally I think we could go one step further. While I like this better than events or inheritance, it still forces you to implement a specific method name. What if that name collides? Furthermore if you have lots of these you end up either with large classes inheriting multiple interfaces to implement one method, or lots of small classes. Also, if you had one class that wanted to manage messages from two different subscribers differently, it wouldn't be able to because the interface can't be overloaded. This brings me to using delegates directly. In general, every time I think about creating an interface for something, and if that interface contains only one method, I start thinking a delegate is a better approach. Now, that said delegates don't accomplish everything an interface can. Obviously having the interface allows you to refer to the classes that implement the interface which can be very handy. In this case, though, really all you want is a method to handle the messages. So let's look at a method delegate: 1: // using delegation via delegate - omitting argument null checks and halt logic 2: public sealed class MessageListener 3: { 4: private ISubscriber _subscriber; 5: private Action<Message> _handler; 6: private bool _isHalted = false; 7: private Thread _messageThread; 8:  9: // assign the subscriber and start the messaging loop 10: public MessageListener(ISubscriber subscriber, Action<Message> handler) 11: { 12: _subscriber = subscriber; 13: _handler = handler; 14: _messageThread = new Thread(MessageLoop); 15: _messageThread.Start(); 16: } 17:  18: // handle the looping in the thread 19: private void MessageLoop() 20: { 21: while(!_isHalted) 22: { 23: // as long as processing, wait 1 second for message 24: Message msg = _subscriber.Receive(TimeSpan.FromSeconds(1)); 25: if(msg != null) 26: { 27: _handler(msg); 28: } 29: } 30: } 31: } Here the MessageListener now takes an Action<Message>.  For those of you unfamiliar with the pre-defined delegate types in .NET, that is a method with the signature: void SomeMethodName(Message). The great thing about delegates is it gives you a lot of power. You could create an anonymous delegate, a lambda, or specify any other method as long as it satisfies the Action<Message> signature. This way, you don't need to define an arbitrary helper class or name the method a specific thing. Incidentally, we could combine both the interface and delegate approach to allow maximum flexibility. Doing this, the user could either pass in a delegate, or specify a delegate interface: 1: // using delegation - give users choice of interface or delegate 2: public sealed class MessageListener 3: { 4: private ISubscriber _subscriber; 5: private Action<Message> _handler; 6: private bool _isHalted = false; 7: private Thread _messageThread; 8:  9: // assign the subscriber and start the messaging loop 10: public MessageListener(ISubscriber subscriber, Action<Message> handler) 11: { 12: _subscriber = subscriber; 13: _handler = handler; 14: _messageThread = new Thread(MessageLoop); 15: _messageThread.Start(); 16: } 17:  18: // passes the interface method as a delegate using method group 19: public MessageListener(ISubscriber subscriber, IMessageHandler handler) 20: : this(subscriber, handler.OnMessageReceived) 21: { 22: } 23:  24: // handle the looping in the thread 25: private void MessageLoop() 26: { 27: while(!_isHalted) 28: { 29: // as long as processing, wait 1 second for message 30: Message msg = _subscriber.Receive(TimeSpan.FromSeconds(1)); 31: if(msg != null) 32: { 33: _handler(msg); 34: } 35: } 36: } 37: } } This is the method I tend to prefer because it allows the user of the class to choose which method works best for them. You may be curious about the actual performance of these different methods. 1: Enter iterations: 2: 1000000 3:  4: Inheritance took 4 ms. 5: Events took 7 ms. 6: Interface delegation took 4 ms. 7: Lambda delegate took 5 ms. Before you get too caught up in the numbers, however, keep in mind that this is performance over over 1,000,000 iterations. Since they are all < 10 ms which boils down to fractions of a micro-second per iteration so really any of them are a fine choice performance wise. As such, I think the choice of what to do really boils down to what you're trying to do. Here's my guidelines: Inheritance should be used only when defining a collection of related types with implementation specific behaviors, it should not be used as a hook for users to add their own functionality. Events should be used when subscription is optional or multi-cast is desired. Interface delegation should be used when you wish to refer to implementing classes by the interface type or if the type requires several methods to be implemented. Delegate method delegation should be used when you only need to provide one method and do not need to refer to implementers by the interface name.

    Read the article

  • Handling Errors In PHP When Using MVC

    - by James Jeffery
    I've been using Codeigniter a lot recently, but one thing that gets on my nerves is handling errors and displaying them to the user. I've never been good at handling errors without it getting messy. My main concern is when returning errors to the user. Is it good practice to use exceptions and throw/catch exceptions rather than returning 0 or 1 from functions and then using if/else to handle the errors. Thus, making it easier to inform the user about the issue. I tend to go away from exceptions. My Java tutor at university some years ago told me "exceptions shouldn't be used in production code it's more for debugging". I get the feeling he was lying. But, an example, I have code that adds a user to a database. During the process more than 1 thing could go wrong, such as a database issue, a duplicate entry, a server issue, etc. When an issue happens during registration the user needs to know about it. What's the best way to handle errors in PHP, keeping in mind that I'm using an MVC framework.

    Read the article

  • How to update a game off a database

    - by James Clifton
    I am currently writing a sports strategy management game (cricket) in PHP, with a MYSQL database, and I have come across one stumbling block - how do I update games where neither player is online? Cricket is a game played between two players, and when they (or one of them) is online then everything is fine; but what if neither player is online? This occurs when championship games are played, and these games need to happen at certain times for game reasons. At the moment I have a private web page that updates every 5 seconds, and each time it loads all games are updated; but then I have the problem that when my private web page stops (for example my computer crashes or my web browser plays up) the game stops updating! Any suggestions?

    Read the article

  • Nginx or Apache for a VPS?

    - by James
    I consider myself to be an inexperienced user/administrator when it comes to running my VPS. I can get by with a few CLI commands, I can set up Webmin and I can set up Yum repos, but beyond the very basic stuff, I'm out of my depth. So far, I'm running Apache. I don't know it particularly well, but I can get by with editing httpd.conf if I'm told what to edit. I've heard good things about Nginx and that it's not as resource-hungry as Apache. I'd like to give it a go, but I can't find any information about its suitability for administrators like me, with little experience of sysadmin or web server config. Webmin now has support for Nginx, so getting it installed and running probably won't be too much of a problem. What I'm wondering is, from a site administrator perspective, is running Nginx as transparent as running Apache? IE, at the moment, I can just throw up Wordpress and Drupal sites without having much to worry about or having to make any config changes to Apache. Would Nginx be as transparent?

    Read the article

  • C#/.NET Little Wonders &ndash; Cross Calling Constructors

    - by James Michael Hare
    Just a small post today, it’s the final iteration before our release and things are crazy here!  This is another little tidbit that I love using, and it should be fairly common knowledge, yet I’ve noticed many times that less experienced developers tend to have redundant constructor code when they overload their constructors. The Problem – repetitive code is less maintainable Let’s say you were designing a messaging system, and so you want to create a class to represent the properties for a Receiver, so perhaps you design a ReceiverProperties class to represent this collection of properties. Perhaps, you decide to make ReceiverProperties immutable, and so you have several constructors that you can use for alternative construction: 1: // Constructs a set of receiver properties. 2: public ReceiverProperties(ReceiverType receiverType, string source, bool isDurable, bool isBuffered) 3: { 4: ReceiverType = receiverType; 5: Source = source; 6: IsDurable = isDurable; 7: IsBuffered = isBuffered; 8: } 9: 10: // Constructs a set of receiver properties with buffering on by default. 11: public ReceiverProperties(ReceiverType receiverType, string source, bool isDurable) 12: { 13: ReceiverType = receiverType; 14: Source = source; 15: IsDurable = isDurable; 16: IsBuffered = true; 17: } 18:  19: // Constructs a set of receiver properties with buffering on and durability off. 20: public ReceiverProperties(ReceiverType receiverType, string source) 21: { 22: ReceiverType = receiverType; 23: Source = source; 24: IsDurable = false; 25: IsBuffered = true; 26: } Note: keep in mind this is just a simple example for illustration, and in same cases default parameters can also help clean this up, but they have issues of their own. While strictly speaking, there is nothing wrong with this code, logically, it suffers from maintainability flaws.  Consider what happens if you add a new property to the class?  You have to remember to guarantee that it is set appropriately in every constructor call. This can cause subtle bugs and becomes even uglier when the constructors do more complex logic, error handling, or there are numerous potential overloads (especially if you can’t easily see them all on one screen’s height). The Solution – cross-calling constructors I’d wager nearly everyone knows how to call your base class’s constructor, but you can also cross-call to one of the constructors in the same class by using the this keyword in the same way you use base to call a base constructor. 1: // Constructs a set of receiver properties. 2: public ReceiverProperties(ReceiverType receiverType, string source, bool isDurable, bool isBuffered) 3: { 4: ReceiverType = receiverType; 5: Source = source; 6: IsDurable = isDurable; 7: IsBuffered = isBuffered; 8: } 9: 10: // Constructs a set of receiver properties with buffering on by default. 11: public ReceiverProperties(ReceiverType receiverType, string source, bool isDurable) 12: : this(receiverType, source, isDurable, true) 13: { 14: } 15:  16: // Constructs a set of receiver properties with buffering on and durability off. 17: public ReceiverProperties(ReceiverType receiverType, string source) 18: : this(receiverType, source, false, true) 19: { 20: } Notice, there is much less code.  In addition, the code you have has no repetitive logic.  You can define the main constructor that takes all arguments, and the remaining constructors with defaults simply cross-call the main constructor, passing in the defaults. Yes, in some cases default parameters can ease some of this for you, but default parameters only work for compile-time constants (null, string and number literals).  For example, if you were creating a TradingDataAdapter that relied on an implementation of ITradingDao which is the data access object to retreive records from the database, you might want two constructors: one that takes an ITradingDao reference, and a default constructor which constructs a specific ITradingDao for ease of use: 1: public TradingDataAdapter(ITradingDao dao) 2: { 3: _tradingDao = dao; 4:  5: // other constructor logic 6: } 7:  8: public TradingDataAdapter() 9: { 10: _tradingDao = new SqlTradingDao(); 11:  12: // same constructor logic as above 13: }   As you can see, this isn’t something we can solve with a default parameter, but we could with cross-calling constructors: 1: public TradingDataAdapter(ITradingDao dao) 2: { 3: _tradingDao = dao; 4:  5: // other constructor logic 6: } 7:  8: public TradingDataAdapter() 9: : this(new SqlTradingDao()) 10: { 11: }   So in cases like this where you have constructors with non compiler-time constant defaults, default parameters can’t help you and cross-calling constructors is one of your best options. Summary When you have just one constructor doing the job of initializing the class, you can consolidate all your logic and error-handling in one place, thus ensuring that your behavior will be consistent across the constructor calls. This makes the code more maintainable and even easier to read.  There will be some cases where cross-calling constructors may be sub-optimal or not possible (if, for example, the overloaded constructors take completely different types and are not just “defaulting” behaviors). You can also use default parameters, of course, but default parameter behavior in a class hierarchy can be problematic (default values are not inherited and in fact can differ) so sometimes multiple constructors are actually preferable. Regardless of why you may need to have multiple constructors, consider cross-calling where you can to reduce redundant logic and clean up the code.   Technorati Tags: C#,.NET,Little Wonders

    Read the article

  • Can anyone recommend online .Net training courses?

    - by james
    I am looking for peoples experiences with paid for online .Net training courses. In your experience, are these an able replacement for in-person training? Are they better than the many free ones provided on MSDN and the like? Are there any specific paid for ones you'd recommend? I usually prefer general book/web research myself, I have one specific provider in mind that looks really good, but I'll omit this for fear of advertising :)

    Read the article

  • WebLogic not reading boot.properties 11.1.1.x

    - by James Taylor
    In WebLogic 11.1.1.1 the boot.properties file was stored in the $MW_HOME/user_projects/domains/[domain] directory. It would be read at startup and there would be no requirement to enter username and password. In later releases the location has changed to $MW_HOME/user_projects/domains/[domain]/servers/[managed_server]/security In most instances you will need to create the security directory If you want to specify a custom directory add the following to the startup scripts for the server. -Dweblogic.system.BootIdentityFile=[loc]/boot.properties create a boot.properties file using the following entry username=<adminuser> password=<password>

    Read the article

  • WebLogic Server–Use the Execution Context ID in Applications–Lessons From Hansel and Gretel

    - by james.bayer
    I learned a neat trick this week.  Don’t let your breadcrumbs go to waste like Hansel and Gretel did!  Keep track of the code path, logs and errors for each request as they flow through the system.  Earlier this week an OTN forum post in the WLS – General category by Oracle Ace John Stegeman asked a question how to retrieve the Execution Context ID so that it could be used on an error page that a user could provide to a help desk or use to check with application administrators so they could look up what went wrong.  What is the Execution Context ID (ECID)?  Fusion Middleware injects an ECID as a request enters the system and it says with the request as it flows from Oracle HTTP Server to Oracle Web Cache to multiple WebLogic Servers to the Oracle Database. It’s a way to uniquely identify a request across tiers.  According to the documentation it’s: The value of the ECID is a unique identifier that can be used to correlate individual events as being part of the same request execution flow. For example, events that are identified as being related to a particular request typically have the same ECID value.  The format of the ECID string itself is determined by an internal mechanism that is subject to change; therefore, you should not have or place any dependencies on that format. The novel idea that I see John had was to extend this concept beyond the diagnostic information that is captured by Fusion Middleware.  Why not also use this identifier in your logs and errors so you can correlate even more information together!  Your logging might already identify the user, so why not identify the request so you filter down even more.  All you need to do inside of WebLogic Server to get ahold of this information is invoke DiagnosticConextHelper: weblogic.diagnostics.context.DiagnosticContextHelper.getContextId() This class has other helpful methods to see other values tracked by the diagnostics framework too.  This way I can see even more detail and get information across tiers. In performance profiling, this can be very handy to track down where time is being spent in code.  I’ve blogged and made videos about this before.  JRockit Flight Recorder can use the WLDF Diagnostic Volume in WLS 10.3.3+ to automatically capture and correlate lots of helpful information for each request without installing any special agents and with the out-of-the-box JRockit and WLS settings!  You can see here how information is displayed in JRockit Flight Recorder about a single request as it calls a Servlet, which calls an EJB, which gets a DB connection, which starts a transaction, etc.  You can get timings around everything and even see the SQL that is used. http://download.oracle.com/docs/cd/E21764_01/web.1111/e13714/using_flightrecorder.htm#WLDFC480 Recent versions of the WLS console also are able to visualize this data too, so it works with other JVMs besides JRockit when you turn on WLDF instrumentation. I wrote a little sample application that verified to myself that the ECID did actually cross JVM boundaries.  I invoked a Servlet in one JVM, which acted as an EJB client to Stateless Session Bean running in another JVM.  Each call returned the same ECID.  You need to turn on WLDF Instrumentation for this to work otherwise the framework returns null.  I’m glad John put me on to this API as I have some interesting ideas on how to correlate some information together.

    Read the article

  • Any valid reason to Nest Master Pages in ASP.Net rather than Inherit?

    - by James P. Wright
    Currently in a debate at work and I cannot fathom why someone would intentionally avoid Inheritance with Master Pages. For reference here is the project setup: BaseProject MainMasterPage SomeEvent SiteProject SiteMasterPage nested MainMasterPage OtherSiteProject MainMasterPage (from BaseProject) The debate came up because some code in BaseProject needs to know about "SomeEvent". With the setup above, the code in BaseProject needs to call this.Master.Master. That same code in BaseProject also applies to OtherSiteProject which is just accessed as this.Master. SiteMasterPage has no code differences, only HTML differences. If SiteMasterPage Inherits MainMasterPage rather than Nests it, then all code is valid as this.Master. Can anyone think of a reason why to use a Nested Master Page here instead of an Inherited one?

    Read the article

  • Perl: Negative look behind regex question [migrated]

    - by James
    The Perlre in Perldoc didn't go into much detail on negative look around but I tried testing it, and didn't work as expected. I want to see if I can differentiate a C preprocessor macro definition (e.g. #define MAX(X) ....) from actual usage (y = MAX(x);), but it didn't work as expected. my $macroName = 'MAX'; my $macroCall = "y = MAX(X);"; my $macroDef = "# define MAX(X)"; my $boundary = qr{\b$macroName\b}; my $bstr = " MAX(X)"; if($bstr =~ /$boundary/) { print "boundary: $bstr matches: $boundary\n"; } else { print "Error: no match: boundary: $bstr, $boundary\n"; } my $negLookBehind = qr{(?<!define)\b$macroName\b}; if($macroCall =~ /$negLookBehind/) # "y = MAX(X)" matches "(?<!define)\bMAX\b" { print "negative look behind: $macroCall matches: $negLookBehind\n"; } else { print "no match: negative look behind: $macroCall, $negLookBehind\n"; } if($macroDef =~ /$negLookBehind/) # "#define MAX(X)" should not match "(?<!define)\bMAX\b" { print "Error: negative look behind: $macroDef matches: $negLookBehind\n"; } else { print "no match: negative look behind: $macroDef, $negLookBehind\n"; } It seems that both $macroDef and $macroCall seem to match regex /(?<!define)\b$macroName\b/. I backed off from the original /(?<\#)\s*(?<!define)\b$macroName\b/ since that didn't work either. So what did I screw up? Also does Perl allow chaining of multiple look around expressions?

    Read the article

  • Are there any actual case studies on rewrites of software success/failure rates?

    - by James Drinkard
    I've seen multiple posts about rewrites of applications being bad, peoples experiences about it here on Programmers, and an article I've ready by Joel Splosky on the subject, but no hard evidence of case studies. Other than the two examples Joel gave and some other posts here, what do you do with a bad codebase and how do you decide what to do with it based on real studies? For the case in point, there are two clients I know of that both have old legacy code. They keep limping along with it because as one of them found out, a rewrite was a disaster, it was expensive and didn't really work to improve the code much. That customer has some very complicated business logic as the rewriters quickly found out. In both cases, these are mission critical applications that brings in a lot of revenue for the company. The one that attempted the rewrite felt that they would hit a brick wall at some point if the legacy software didn't get upgraded at some point in the future. To me, that kind of risk warrants research and analysis to ensure a successful path. My question is have there been actual case studies that have investigated this? I wouldn't want to attempt a major rewrite without knowing some best practices, pitfalls, and successes based on actual studies. Aftermath: okay, I was wrong, I did find one article: Rewrite or Reuse. They did a study on a Cobol app that was converted to Java.

    Read the article

< Previous Page | 8 9 10 11 12 13 14 15 16 17 18 19  | Next Page >