Scheme in the real world
10 comments
We're using Scheme in our startup and I recommend it. Our web app is built in Kawa and runs on the Tomcat server. We had been using the PLT web server but left over some concerns with the way it was using memory. They've fixed that and now I think switching was a mistake - we should have just stuck with PLT. Kawa... hoo boy. It's got some issues. It's still Scheme, though, mostly.
So for the web I would use PLT and its great web server. In fact I do use it for a little internal app and love being able to write a web app in the 'direct style' ie just write a web page interaction with a user as a function execution. You can just assign a lambda to a link, saying 'when the user clicks this link, execute this lambda'. It's great. However, that links won't be bookmarkable nor indexable by search engines, so if some of your pages need to be, you have to make the links to them be traditional links. So the continuations-based stuff lends itself to the parts of your app that are private to a single user (like taking user input or showing them their email) and/or shouldn't show up in a search engine anyway. This would include anything that you would describe as a 'wizard'.
NASA (or is it just the Air Force?) is using PLT Scheme to control a large telescope in New Mexico (http://news.ycombinator.com/item?id=75219) to great success. So for soft real-time stuff you should be good. Just from using it, I wouldn't recommend the default garbage collector with PLT for anything like hard real-time stuff. But there are other GCs, so ask around. The PLT mailing list is excellent: http://www.plt-scheme.org/maillist/
So for the web I would use PLT and its great web server. In fact I do use it for a little internal app and love being able to write a web app in the 'direct style' ie just write a web page interaction with a user as a function execution. You can just assign a lambda to a link, saying 'when the user clicks this link, execute this lambda'. It's great. However, that links won't be bookmarkable nor indexable by search engines, so if some of your pages need to be, you have to make the links to them be traditional links. So the continuations-based stuff lends itself to the parts of your app that are private to a single user (like taking user input or showing them their email) and/or shouldn't show up in a search engine anyway. This would include anything that you would describe as a 'wizard'.
NASA (or is it just the Air Force?) is using PLT Scheme to control a large telescope in New Mexico (http://news.ycombinator.com/item?id=75219) to great success. So for soft real-time stuff you should be good. Just from using it, I wouldn't recommend the default garbage collector with PLT for anything like hard real-time stuff. But there are other GCs, so ask around. The PLT mailing list is excellent: http://www.plt-scheme.org/maillist/
Good luck with your startup, make sure you tell us when it launches !
Thanks! We'll definitely post our launch here.
Zak, why does it matter to you if someone is using Scheme in the real world? The question you should be asking is if you can use it to your advantage to outperform others who aren't using it.
So far, the biggest issue with Common Lisp I ran into while writing Weblocks is lack of continuations. There were other issues but they were minor annoyances that were fairly easy to overcome. Right now I'm working on a delimited continuations library for CL. Writing one really isn't as bad as it sounds - I think On Lisp unnecessarily scares people from rolling their own code walker.
Out of curiosity, other than continuations, what advantage do you see in using Scheme over CL?
So far, the biggest issue with Common Lisp I ran into while writing Weblocks is lack of continuations. There were other issues but they were minor annoyances that were fairly easy to overcome. Right now I'm working on a delimited continuations library for CL. Writing one really isn't as bad as it sounds - I think On Lisp unnecessarily scares people from rolling their own code walker.
Out of curiosity, other than continuations, what advantage do you see in using Scheme over CL?
I really hate it when someone answers a question with "your question doesn't matter. What you should be asking is [whatever I think you should be asking] "
I assume the person asking the question has good reasons for asking it. If I can I answer. If not, I don't.
Coming to the question proper, there are often good reasons to ask if other people use / don't use a particular technology.
To answer, I use a custom variant of scheme to write machine- learning systems on a cluster of cots machines. It compiles down to optimized C, but I write in scheme.
I assume the person asking the question has good reasons for asking it. If I can I answer. If not, I don't.
Coming to the question proper, there are often good reasons to ask if other people use / don't use a particular technology.
To answer, I use a custom variant of scheme to write machine- learning systems on a cluster of cots machines. It compiles down to optimized C, but I write in scheme.
A big part of it is the fact that Scheme is simple and Common Lisp is not. This is a "soft" feature, but I think it's important. I want to try out Scheme and see if it feels better.
I ask because I'd like to get a perspective on the tools that are available from people who actually make things with them. The programming language I use might be a "secret weapon" in some situations, but most of the people it would give me an advantage over wouldn't even understand why it's an advantage.
I ask because I'd like to get a perspective on the tools that are available from people who actually make things with them. The programming language I use might be a "secret weapon" in some situations, but most of the people it would give me an advantage over wouldn't even understand why it's an advantage.
The problem I had with Scheme was that it doesn't stay simple. The core is certainly more elegant than Common Lisp's, but all the cruft has a nasty way of sneaking back in through SRFIs and big hairy macros.
I think continuations is an advantage if you don't use ajax.
When you use Ajax, continuations in a web framework aren't needed. This has been my experience with the last two web applications I've written using CL and UCW.
Good timing Scheme is undergoing quite a bit of change that should culminate in about 6 months.
I have used scheme at work and for the last 6 weeks on a private project in earnest.
One issues is that there are a lot of scheme implementations out there. Some good, great, average, and others bit rotted. Only mention two here.
DrScheme/MzScheme is pretty amazing. It is fully loaded and few know of its full capabilities. The fact that PG can just "write" Arc in MzScheme and some other pretty neat things that are doable in MzScheme is not an accident. DrScheme/MzScheme is much, much more then Scheme in a way not too many people understand.
The second implementation is Larceny. True native compiler. Plenty fast. The compiler is written in Scheme (small C runtime). Its hackable from top to bottom. Extensible and did I say fast? Threading is fair or green threads using continuations/call-cc (no O/S threads). Don't like the Scheduler? Modify it, its in tasking.sch The interrupt handler is a lambda function, just redefine it. Don't like the i386 code output, hack the assembler or the optimized 5 pass compiler. Its all in Scheme.
Don't like standard poll, I added epoll via the FFI in a few hours.
Larceny has been around for years, but is undergoing a resurgence to bring it up "closer" to the new R6RS standard.
I think it has great potential. Be advised that do to the rapid changes in development underway that there are some rough edges in certain "library" areas, however, the core compiler is solid and stable.
Awhile back Slime was ported over to Scheme48 (not a system for general real world use, but a fun hacking implementation). The compiler group should have Larceny stable with many of the libraries and features of the new R6RS standard (some will not implemented). IF SLIME scheme backend was ported over to Larceny as well. It could be the start of something special.
I have used scheme at work and for the last 6 weeks on a private project in earnest.
One issues is that there are a lot of scheme implementations out there. Some good, great, average, and others bit rotted. Only mention two here.
DrScheme/MzScheme is pretty amazing. It is fully loaded and few know of its full capabilities. The fact that PG can just "write" Arc in MzScheme and some other pretty neat things that are doable in MzScheme is not an accident. DrScheme/MzScheme is much, much more then Scheme in a way not too many people understand.
The second implementation is Larceny. True native compiler. Plenty fast. The compiler is written in Scheme (small C runtime). Its hackable from top to bottom. Extensible and did I say fast? Threading is fair or green threads using continuations/call-cc (no O/S threads). Don't like the Scheduler? Modify it, its in tasking.sch The interrupt handler is a lambda function, just redefine it. Don't like the i386 code output, hack the assembler or the optimized 5 pass compiler. Its all in Scheme.
Don't like standard poll, I added epoll via the FFI in a few hours.
Larceny has been around for years, but is undergoing a resurgence to bring it up "closer" to the new R6RS standard.
I think it has great potential. Be advised that do to the rapid changes in development underway that there are some rough edges in certain "library" areas, however, the core compiler is solid and stable.
Awhile back Slime was ported over to Scheme48 (not a system for general real world use, but a fun hacking implementation). The compiler group should have Larceny stable with many of the libraries and features of the new R6RS standard (some will not implemented). IF SLIME scheme backend was ported over to Larceny as well. It could be the start of something special.
You used it to post this question. News.YC is written in a version of Arc that runs on MzScheme.
yeah that is really really cool. Any hints on Lines Of Code
Don't tell if you don't want to, or if you think its too long ! Maybe paren count is more appropriate ;)
I am guessing amazingly concise.
Don't tell if you don't want to, or if you think its too long ! Maybe paren count is more appropriate ;)
I am guessing amazingly concise.
Ask him to run trav (tree-fold) on the code: http://news.ycombinator.com/item?id=32766
Currently the code tree is 9573 nodes. About
1500 lines of code. It's fairly tight at this stage,
because I recently went through it and squeezed out a
lot of the fat.
wow! 9573 nodes is very concise!
I just ran a scheme equivalent of trav on one of my code files and get a count of 300 (on a very small file). This could of course mean that my code is bloated, but I am impressed anyway!
Just curious, but does arc have its own macro system? Or do you use MZScheme's macros?
I just ran a scheme equivalent of trav on one of my code files and get a count of 300 (on a very small file). This could of course mean that my code is bloated, but I am impressed anyway!
Just curious, but does arc have its own macro system? Or do you use MZScheme's macros?
I'm using Scheme both in the real world of my daytime employment, and in the fake world of my personal project. It was web development that brought me back to Scheme after 12 years of C. I wrote BRL in order to make server-side web development easier, especially when making database-driven sites.
What I don't understand is why you want something "a little less bulky"? Can't you use a subset of CL?
What I don't understand is why you want something "a little less bulky"? Can't you use a subset of CL?
I think everybody who uses CL uses a subset. The problem is that everyone uses a different subset, which can be an issue when using libraries.
Do you mind sharing where you work?
Like most people I know of who use Scheme or other Lisps in the real world, I'm in the financial services industry. Here's a job opening for someone who would report to me:
https://tbe.taleo.net/NA4/ats/careers/requisition.jsp?org=EA...
https://tbe.taleo.net/NA4/ats/careers/requisition.jsp?org=EA...
So, IIRC DMX runs at about 44hz if you've got a full dimmer pack (which isn't hard with moving lights since they each take a bunch of channels). That means you get about 20ms to compute each cycle, which strikes me as pretty soft indeed if you're running on PC hardware... I wouldn't worry about performance until you see problems.
Eons ago, I remember building a DMX controller by ripping out the clock from a standard RS485 card and soldering in a new (8khz?) one from Radio Shack so the baud rate would be right...
Eons ago, I remember building a DMX controller by ripping out the clock from a standard RS485 card and soldering in a new (8khz?) one from Radio Shack so the baud rate would be right...
I'm not worrying too much about performance yet, but I do see it as a potential issue. I'm trying to design a scriptable effects engine that could allow users to do things that are inefficient. This may sound like premature optimization, but I don't really see a downside to preferring implementations that tend to be fast.
You should try hop - best Scheme for the web today:
http://hop.inria.fr/ http://hop.inria.fr/usr/local/share/hop/weblets/home/article...
Also, it has a scheme-to-javascript compiler inside it...
If you really like Java, SISCWeb is also nice.
Also, it has a scheme-to-javascript compiler inside it...
If you really like Java, SISCWeb is also nice.
Recently I did a consulting job for a project of a kind of network backbone testing suites used by one of major telcom providers in Japan. The suite was mainly in CL, but my part was an independent, stand-alone tool and I used Scheme (Gauche). It was a distributed, multithreaded app that opened raw sockets and sent/received fair amount of ICMP packets. The main app was just around 1000 LOC in Scheme, though it used quite a few libraries came with Gauche.
Speed didn't become an issue even Gauche wasn't a native compiler (it compiles to VM), for the network driver was always the bottleneck. Yet I did care not to cons in the inner loop.
Speed didn't become an issue even Gauche wasn't a native compiler (it compiles to VM), for the network driver was always the bottleneck. Yet I did care not to cons in the inner loop.
Out of curiosity, why would you want to use a garbage collected language at all if you have real time requirements.
I don't want to do memory management. It's the sort of error-prone process I feel is generally better left to a machine. I think the timing uncertainties inherent in GC can be mitigated to the point that they don't affect performance.
You are right. RScheme has a real time garbage collector
http://en.wikipedia.org/wiki/RScheme
http://en.wikipedia.org/wiki/RScheme
I certainly hope so (that there are people using it). And increasing it.
The first is standard webapp development. I've been using Hunchentoot, CL-WHO and friends here, and I'm experimenting with Weblocks. I generally like these tools, but I'm open to trying tools that work differently. I'd like to know what tools the Scheme users here prefer for web work.
The other project is a desktop/embedded (fairly fast PC hardware) system with soft real-time requirements (for the curious, it's a DMX lighting controller). I'm looking for speed, multithreading and some predictability in the garbage collector.