Some notes on a high-performance Python application.
I run SiteTruth (sitetruth.com), which rates web sites for legitimacy, based on what information it can find out about the business behind the web site. I'm going to describe here how the machinery behind this is organized, because I had to solve some problems in Python that I haven't seen solved before. The site is intended mainly to support AJAX applications which query the site for every ad they see. You can download the AdRater client (http://www.sitetruth.com/downloads/adrater.html;) and use the site, if you like. It's an extension for Firefox, written in Javascript. For every web page you visit, it looks for URLs that link to ad sites, and queries the server for a rating, then puts up icons on top of each ad indicating the rating of the advertiser. The client makes the query by sending a URL to an .fcgi program in Python, and gets XML back. So that's the interface. At the server end, there's an Linux/Apache/mod_fcgi/Python server. Requests come in via FCGI, and are assigned to an FCGI server process by Apache. The initial processing is straightforward; there's a MySQL database and a table of domains and ratings. If the site is known, a rating is returned immediately. This is all standard FCGI. If the domain hasn't been rated yet, things get interesting. The server returns an XML reply with a status code that tells the client to display a busy icon and retry in five seconds. Then the process of rating a site has to be started. This takes more resources and needs from 15 seconds to a minute, as pages from the site are read and processed. So we don't want to do rating inside the FCGI processes. We want FCGI processing to remain fast even during periods of heavy rating load. And we may need to spread the processing over multiple computers. So the FCGI program puts a rating request into the database, in a MySQL table of type ENGINE=MEMORY. This is just an in-memory table, something that MySQL supports but isn't used much. Each rating server has a rating scheduler process, which repeatedly reads from that table, looking for work to do. When it finds work, it marks the task as in process. The rating scheduler launches multiple subprocesses to do ratings, all of which run at a lower priority than the rest of the system. The rating scheduler communicates with its subprocesses via pipes and Pickle. Launching a new subprocess for each rating is too slow; it adds several seconds as CPython loads code and starts up. So the subprocesses are reusable, like FCGI tasks. Every 100 uses or so, we terminate each subprocess and start another one, in case of memory leaks. (There seems to be a leak we can't find in M2Crypto. Guido couldn't find it either when he used M2Crypto, as he wrote in his blog.) Each rating process only rates one site at a time, but is multithreaded so it can read multiple pages from the site, and other remote data sources like BBBonline, at one time. This allows us to get a rating within 15 seconds or so. When the site is rated, the database is updated, and the next request back at the FCGI program level will return the rating. We won't have to look at that domain for another month. The system components can run on multiple machines. One can add rating capacity by adding another rating server and pointing it at the same database. FCGI capacity can be added by adding more FCGI servers and a load balancer. Adding database capacity is harder, because that means going to MySQL replication, which creates coordination problems we haven't dealt with yet. Also, since multiple processes are running on each CPU, multicore CPUs help. Using MySQL as a queueing engine across multiple servers is unusual, but it works well. It has the nice feature that the queue ordering can be anything you can write in a SELECT statement. So we put fair queueing in the rating scheduler; multiple requests from the same IP address compete with each other, not with those from other IP addresses. So no one site can use up all the rating capacity. Another useful property of using MySQL for coordination is that we can have internal web pages that make queries and display the system and queue status. This is easy to do from the outside when the queues are in MySQL. It's tough to do that when they're inside some process. We log errors in a database table, not text files, for the same reason. In addition to specific problem logging, all programs have a final try block around the whole program that does a stack backtrace and puts that in a log entry in MySQL. All servers log to the same database. Looking at this architecture, it was put together from off the shelf parts, but not the parts that have big fan bases. FCGI isn't used much. The MySQL memory engine isn't used much. MySQL advisory locking (SELECT GET LOCK(lockname,timeout)) isn't used much. Pickle isn't used much over pipes. M2Crypto isn't used much. We've spent much time finding
Re: Some notes on a high-performance Python application.
Am Mittwoch, 26. März 2008 17:33:43 schrieb John Nagle: ... Using MySQL as a queueing engine across multiple servers is unusual, but it works well. It has the nice feature that the queue ordering can be anything you can write in a SELECT statement. So we put fair queueing in the rating scheduler; multiple requests from the same IP address compete with each other, not with those from other IP addresses. So no one site can use up all the rating capacity. ... Does anyone else architect their systems like this? A Xen(tm) management system I've written at least shares this aspect in that the RPC subsystem for communication between the frontend and the backends is basically a (MySQL) database table which is regularily queried by all backends that work on VHosts to change the state (in the form of a command) according to what the user specifies in the (Web-)UI. FWIW, the system is based on SQLObject and CherryPy, doing most of the parallel tasks threaded from a main process (because the largest part of the backends is dealing with I/O from subprocesses [waiting for them to complete]), which is different from what you do. CherryPy is also deployed with the threading server. -- Heiko Wundram -- http://mail.python.org/mailman/listinfo/python-list
Re: Some notes on a high-performance Python application.
Heiko Wundram wrote: Am Mittwoch, 26. März 2008 17:33:43 schrieb John Nagle: ... Using MySQL as a queueing engine across multiple servers is unusual, but it works well. It has the nice feature that the queue ordering can be anything you can write in a SELECT statement. So we put fair queueing in the rating scheduler; multiple requests from the same IP address compete with each other, not with those from other IP addresses. So no one site can use up all the rating capacity. ... Does anyone else architect their systems like this? A Xen(tm) management system I've written at least shares this aspect in that the RPC subsystem for communication between the frontend and the backends is basically a (MySQL) database table which is regularily queried by all backends that work on VHosts to change the state (in the form of a command) according to what the user specifies in the (Web-)UI. I see nothing unusual with this: I vaguely remember that this database approach was teached at my former university as a basic mechanism for distributed systems at least since 1992, but I'd guess much longer... And in one of my projects a RDBMS-based queue was used for a PKI registration server (e.g. for handling the outbound CMP queue). IIRC Microsoft's Biztalk Server also stores inbound and outbound queues in its internal MS-SQL database (which then can be the bottleneck). Ciao, Michael. -- http://mail.python.org/mailman/listinfo/python-list
Re: Some notes on a high-performance Python application.
Am Mittwoch, 26. März 2008 18:54:29 schrieb Michael Ströder: Heiko Wundram wrote: Am Mittwoch, 26. März 2008 17:33:43 schrieb John Nagle: ... Using MySQL as a queueing engine across multiple servers is unusual, but it works well. It has the nice feature that the queue ordering can be anything you can write in a SELECT statement. So we put fair queueing in the rating scheduler; multiple requests from the same IP address compete with each other, not with those from other IP addresses. So no one site can use up all the rating capacity. ... Does anyone else architect their systems like this? A Xen(tm) management system I've written at least shares this aspect in that the RPC subsystem for communication between the frontend and the backends is basically a (MySQL) database table which is regularily queried by all backends that work on VHosts to change the state (in the form of a command) according to what the user specifies in the (Web-)UI. I vaguely remember that this database approach was teached at my former university as a basic mechanism for distributed systems at least since 1992, but I'd guess much longer... I didn't say it was unusual or frowned upon (and I was also taught this at uni IIRC as a means to easily distribute systems which don't have specific requirements for response time to RPC requests), but anyway, as you noted for Biztalk, it's much easier to hit bottlenecks with a polling-style RPC than with a true RPC system, as I've come to experience when the number of nodes (i.e., backends) grew over the last year and a half. That's what's basically causing a re-consideration to move from DB-style RPC to socket-based RPC, which is going to happen at some point in time for the system noted above (but I've sinced changed jobs and am now only a consulting developer for that anyway, so it won't be my job to do the dirty migration and the redesign ;-)). -- Heiko Wundram -- http://mail.python.org/mailman/listinfo/python-list
Re: Some notes on a high-performance Python application.
Heiko Wundram wrote: Am Mittwoch, 26. März 2008 18:54:29 schrieb Michael Ströder: Heiko Wundram wrote: Am Mittwoch, 26. März 2008 17:33:43 schrieb John Nagle: I didn't say it was unusual or frowned upon (and I was also taught this at uni IIRC as a means to easily distribute systems which don't have specific requirements for response time to RPC requests), but anyway, as you noted for Biztalk, it's much easier to hit bottlenecks with a polling-style RPC than with a true RPC system, as I've come to experience when the number of nodes (i.e., backends) grew over the last year and a half. I know, I don't like the polling either. The time scale is such that the poll delay isn't a problem, though, and because it's using the MySQL MEMORY engine, there's no disk I/O. After completing a request, the rating scheduler immediately queries the database, so there's no lost time if there's a queue. The polling delay only applies when a rating server is idle. I miss QNX, which has good message passing primitives. Linux is weak in that area. John Nagle -- http://mail.python.org/mailman/listinfo/python-list