5 ms·
I am a former Archive employee. I can't speak to their current infrastructure (though more of it is open source now - http://archive-access.sourceforge.net/pro
by mmagin 12y ago
I am a former Archive employee. I can't speak to their current infrastructure (though more of it is open source now - http://archive-access.sourceforge.net/projects/wayback/ http://archive-access.sourceforge.net/projects/wayback/ ), but as far as the wayback machine, there was no SQL database anywhere in it. For the purposes of making the wayback machine go:
- Archived data was in ARC file format (predecessor to http://en.wikipedia.org/wiki/Web_ARChive http://en.wikipedia.org/wiki/Web_ARChive) which is essentially a concatenation of seperately gzipped records. That is, you can seek to a particular offset and start decompressing a record. Thus you could get at any archived web page with a triple (server, filename, file-offset) Thus it was spread across a lot of commodity grade machines.
- An sorted index of all the content was built that would let you lookup (url) and give you a list of times or (url, time) to (filename, file-offset). It was implemented by building a sorted text file (first sorted on the url, second on the time) and sharded across many machines by simply splitting it into N roughly equal sizes. Binary search across a sorted text file is surprisingly fast -- in part because the first few points you look at in the file remain cached in RAM, since you hit them frequently.
- (Here's where I'm a little rusty) The web frontend would get a request, query the appropriate index machine. Then it would use a little mechanism (network broadcast maybe?) to find out what server that (unique) filename was on, then it would request the particular record from that server.
(Edit: FYI, my knowledge is 5 years old now. I know they've done some things to keep the index more current than they did back then.)
At the very least, I'd think about getting your blobs out of MySQL and putting them in the filesystem. Filesystems are good at this stuff. You can certainly do something as simple as a SHA-1 hash of the content as the filename, and then depending on your filesystem's performance characteristics, you can have a couple levels in the tree you store them in. e.g. da39a3ee5e6b4b0d3255bfef95601890afd80709 goes into the directory da/39/ Then you stick da39a3ee5e6b4b0d3255bfef95601890afd80709 into the 'pointer' field in your table that replaces the actual data. Obviously this design assumes the content of _that_ file doesn't change. If you want to change the data for that row in the table, you have to write a new file in the filesystem and update the 'pointer'.
- rajbot 12y agoThanks! We were writing up a response at the same time: The Wayback Machine data is stored in WARC or ARC files[0] which are written at web crawl time by the Heritrix crawler[1] (or other crawlers) and stored as regular files in the archive.org storage cluster. Playback is accomplished by binary searching a 2-level index of pointers into the WARC data. The second level of this index is a 20TB compressed sorted list of (url, date, pointer) tuples called CDX records[2]. The first level fits in core, and is a 13GB sorted list of every 3000th entry in the CDX index, with a pointer to larger CDX block. Index lookup works by binary searching the first level list stored in core, then HTTP range-request loading the appropriate second-level blocks from the CDX index. Finally, web page data is loaded by range-requesting WARC data pointed to by the CDX records. Before final output, link re-writing and other transforms are applied to make playback work correctly in the browser. The server stack: - frontend: Tengine + HAProxy to a pool of Wayback tomcat app servers[3] - backend: The redis-backed archive.org metadata API[4] for object location and nginx on linux (via ext4) for data service [0] http://en.wikipedia.org/wiki/Web_ARChive [1] https://github.com/internetarchive/heritrix3 [2] https://github.com/internetarchive/CDX-Writer [3] https://github.com/internetarchive/wayback [4] http://blog.archive.org/2013/07/04/metadata-api/ -sam and raj, Internet Archive
- mmagin 12y agoAnd despite the size, it's so much faster than it used to be! Good work.
- sytelus 12y agoWhy not use hashtable instead of binary search? I'm assuming your index is immutable and query for the data structure is essentially random. Another advantage of looking up item by URL hash may be that you can use hash-prefix to direct the query to appropriate machine (so basically, your infrastructure simply looks like giant distributed hashtable top to bottom with no binary searches required).
- bglazer 12y agoPerhaps caching reasons? If people follow links internally within a domain then randomly distributed hashes would mean a visit to randomly distributed data servers to retrieve every new page. With their architecture the 'CDX' block should contain similar URLs and accessing the linked URLs would could be a seek within the already retrieved block. Just my guess.