• My adventures in Crystal land continue so far without the crushing realisation that it all, in fact, sucks.  Uploading binaries sucks because I'm still stuck on ADSL, but I'll get Crystal installed on my dev server shortly to sort that out.
  
  I wrote a little test app to compare performance of HTTP via the Amber framework, raw HTTP by the bundled server, TCP sockets, UDP sockets, web sockets, and Unix sockets.  Over a Unix socket with a single client and server process on my dev server I can get 50,000 requests per second.  That's mostly consumed by the kernel and not by the app itself, because the packets I'm sending are small and I'm just echoing them back again without any real processing load.
  
  Despite the fact that Crystal is statically typed and compiled, I can write code like this, compile it and run it without a single warning:
  
  
  def handle_client(client)

    while message = client.gets

        client.puts message

    end

  end

  
  

  def serve_tcp

    server = TCPServer.new("localhost", 3000)

    while client = server.accept?

      spawn handle_client(client)

    end

  end

  
  

  def serve_unix

    server = UNIXServer.new("/tmp/amble.sock")

    while client = server.accept?

      spawn handle_client(client)

    end

  end

  
  So TCP and Unix socket servers can share the same handler code transparently, and I haven't needed to define the type of anything anywhere.  It could easily be Ruby code, except that it's fast.
  
  The one notable limitation of Crystal right now is that it's single-threaded.  Python and Ruby are multi-threaded but bound by something called a Global Interpreter Lock (the infamous GIL), which means that only one thread can actually be executing Python or Ruby instructions at any time.  This actually makes things faster because otherwise you have to have locks controlling access to datastructures shared by the threads.
  
  Crystal (currently) just has one operating system thread (plus a second one for the garbage collector) and uses green threads a.k.a fibers for internal concurrency.  You can see that in the above code - every time a connection comes in to the socket I spawn a new thread.  Fibers only use 4K of RAM each so you can have trillions of them if you want.
  
  Internally it's using an event loop, but unlike a trainwreck like Node.js you don't care about that.  You spawn your little fibers and get on with your life.  Got 10,000 concurrent connections?  Who cares?  That's 40MB of RAM; it costs about 20¢ per month.  If you even stopped for a minute to think about it you just wasted more money than it is likely to cost you.
  
  And given that it's ten times faster than Python or Ruby for many tasks, it's like being given a multi-threaded language and a 10-core server for free.  Even so, improved concurrency is at the top of the development roadmap.
  
  I'm liking Crystal a lot so far, with the caveat that library support is (unsurprisingly) far behind Ruby, let alone Python.  Python has a library for everything.  Except VelocyPack.
  
  Update: Ooh, a brand new LMDB binding!  LMDB is perfect here because it supports multiple independent processes directly mapping the same database.  I did find a couple of earlier bindings but the projects had been dormant a couple of years while the Crystal language has been updated regularly.
  
  LMDB is kind of magical.  It's pretty good at handling concurrent writes (it's properly ACIDic), though it doesn't deliver the transaction performance of high-end database engines.  But being memory-mapped and copy-free it is phenomenally fast at reads.  I ran some tests in Python a while back reading 1GB per second on a small VPS.
  
  With LMDB it doesn't matter that your language has a GIL or uses green threads, because you can just run multiple processes.  It handles transactions and crash recovery for you, the database is a single file, and it comes with a hot backup utility where the backups are themselves LMDB databases.  That is, restore time once you have a backup is zero.
  
  This makes a project I've wanted to do for years suddenly very, very feasible.
  
  
• TSMC has started risk production of 5nm devices.  (WikiChip Fuse)
  
  This extends the use of EUV throughout the production process and significantly shrinks the geometry, packing about 80% more transistors per square millimetre.  Performance and power improvements are expected to be smaller, on the order of 
  10%.
  
  
• But MongoDB is web scale.
  
  A good roundup of terrible trends in web development.
  
  
• Browsers are increasingly disabling disabling click tracking.  (Bleeping Computer)
  
  Except for Firefox among the major players.  Brave also still allows this, and it's likely that MoonPie and Vivaldi and other niche browsers do too.
  
  I didn't know that the ping attribute was at thing.  That's really useful and much better than using JavaScript to achieve the same end, but users should be able to turn it off.
  
  
• Does Apple's "Magic Keyboard" still suck?  Yes.  (ZDNet)

Bonus Video of the Day

Posted by: Rick C at Monday, April 08 2019 02:54 AM (Iwkd4)

Posted by: Pixy Misa at Monday, April 08 2019 04:05 AM (PiXy!)

Posted by: Mauser at Monday, April 08 2019 10:12 AM (Ix1l6)

Posted by: The Brickmuppet at Monday, April 08 2019 11:59 AM (xOgT9)

Posted by: Pixy Misa at Monday, April 08 2019 02:07 PM (PiXy!)

Posted by: Mauser at Tuesday, April 09 2019 01:14 PM (Ix1l6)

Posted by: Pixy Misa at Tuesday, April 09 2019 11:40 PM (PiXy!)

Hide Comments | Add Comment