On Tue, Feb 13, 2018 at 4:52 AM, Igor Gnatenko <ignatenkobr...@fedoraproject.org> wrote: > -----BEGIN PGP SIGNED MESSAGE----- > Hash: SHA256 > > CCing rpm-ecosystem@ ML since it's main location where this message should > have > went 😉 > > On Mon, 2018-02-12 at 23:53 +0200, Jonathan Dieter wrote: >> <tl;dr> >> I've come up with a method of splitting repodata into chunks that can >> be downloaded and combined with chunks that are already on the local >> system to create a byte-for-byte copy of the compressed repodata. >> Tools and scripts are at: >> https://www.jdieter.net/downloads/ >> </tl;dr> >> >> Background: >> With DNF, we're currently downloading ~20MB of repository data every >> time the updates repository changes. >> >> When casync was released, I wondered if we could use it to only >> download the deltas for the repodata. At Flock last summer, I ran some >> tests against the uncompressed repodata and saw a reduction of 30-40% >> from one day to the next, which seemed low, but was a good starting >> point. >> >> Unfortunately, due to the way casync separates each file into thousands >> of compressed chunks, building each file required thousands of (serial) >> downloads which, even on a decent internet connection, took *forever*. >> >> When I talked through the idea with Kevin and Patrick, they also >> pointed out that our mirrors might not be too keen on the idea of >> adding thousands of tiny files that change every day. >> >> >> The Solution(?): >> One potential solution to the "multitude of files" problem is to merge >> the chunks back into a single file, and use HTTP ranges to only >> download the parts of the file we want. An added bonus is that most >> web servers are configured to support hundreds of ranges in one >> request, which greatly reduces the number of requests we have to make. >> >> The other problem with casync is that it's chunk separation is naïve, >> which is why we were only achieving 30-40% savings. But we know what >> the XML file is supposed to look like, so we can separate the chunks on >> the tag boundaries in the XML. >> >> So I've ditched casync altogether and put together a proof-of-concept >> (tentatively named zchunk) that takes an XML file, compresses each tag >> separately, and then concatenates all of them into one file. The tool >> also creates an index file that tells you the sha256sum for each >> compressed chunk and the location of the chunk in the file. >> >> I've also written a small script that will download a zchunk off the >> internet. If you don't specify an old file, it will just download >> everything, but if you specify an old file, it will download the index >> of the new file and compare the sha256sums of each chunk. Any >> checksums that match will be taken from the old file, and the rest will >> be downloaded. >> >> In testing, I've seen savings ranging from 10% (December 17 to today) >> to 95% (yesterday to today). >>
This is fantastic. It'd be really helpful for users in Mageia where its more common for users to have limited bandwidth. >> >> Remaining problems: >> * Zchunk files are bigger than their gzip equivalents. This ranges >> from 5% larger for filelists.xml to 300% larger for primary.xml. >> This can be greatly reduced by chunking primary.xml based on srpm >> rather than rpm, which brings the size increase for primary.xml down >> to roughly 30%. >> >> * Many changes to the metadata can mean a large number of ranges >> requested. I ran a check on our mirrors, and three (out of around >> 150 that had the file I was testing) don't honor range requests at >> all, and three others only honor a small number in a single request. >> A further seven didn't respond at all (not sure if that had >> anything to do with the range requests), and the rest supported >> between 256 and 512 ranges in a single request. We can reduce the >> number of ranges requested by always ordering our packages by date. >> This would ensure that new packages are grouped at the end of the >> xml where they will be grabbed in one contiguous range. > > This would "break" DNF, because libsolv is assigning Id's by the order of > packages in metadata. So if something requires "webserver" and there is > "nginx" > and "httpd" providing it (without versions), then lowest Id is picked up (not > going into details of this). Which means depending on when last update for one > or other was submitted, users will get different results. This is unacceptable > from my POV. > Even without this optimization, we'd still have some savings.. >> * Zchunk files use zlib (it gives better compression than xz with such >> small chunks), but, because they use a custom zdict, they are not gz >> files. This means that we'll need new tools to read and write them. >> (And I am volunteering to do the work here) > > What about zstd? Also in latest version of lz4 there is support for > dictionaries too. > I would suggest using zstd or lz4 too. I already had filed an RFE for createrepo_c to support zstd[1], but if lz4 would be better in this regard, that route is fine too. [1]: https://github.com/rpm-software-management/createrepo_c/issues/82 >> The tools: >> The proof-of-concept tools are all sitting in >> https://www.jdieter.net/downloads/zchunk-scripts/ >> >> They are full of ugly hacks, especially when it comes to parsing the >> XML, there's little to no error reporting, and I didn't comment them >> well at all, but they should work. >> >> If all you want to do is download zchunks, you need to run dl_zchunk.py >> with the url you want to download (ending in .zck) as the first >> parameter. Repodata for various days over the last few weeks is at: >> https://www.jdieter.net/downloads/zchunk-test/ You may need to hover >> over the links to see which is which. The downloads directory is also >> available over rsync at rsync://jdieter.net/downloads/zchunk-test. >> >> dl_zchunk.py doesn't show anything if you download the full file, but >> if you run the command with an old file as the second parameter, it >> will show four numbers: bytes taken from the old file, bytes downloaded >> from the new, total downloaded bytes and total uploaded bytes. >> >> zchunk.py creates a .zck file. To group chunks by source rpm in >> primary.xml, run >> ./zchunk.py <file> rpm:sourcerpm >> >> unzchunk.py decompresses a .zck file to stdout >> >> I realize that there's a lot to digest here, and it's late, so I know I >> missed something. Please let me know if you have any suggestions, >> criticisms or flames, though it might be a few hours before I respond. > > As being someone who tried to work on this problem I very appreciate what you > have done here. We've started with using zsync and results were quite good, > but > zsync is dead and has ton of bugs. Also it requires archives to be ` > - --rsyncable`. So my question is why not to add idx file as additional one > for > existing files instead of inventing new format? The problem is that we will > have to distribute in old format too (for compatibility reasons). > > I'm not sure if trying to do optimizations by XML tags is very good idea > especially because I hope that in future we would stop distributing XML's and > start distributing solv/solvx. solv/solvx form is unstable and internal to libsolv, it's a really bad idea to distribute that. However, I like that this is distributing XML, because then it's relatively easy to implement for tools like changelog browser tools and things like that. One of the things I'm resigning myself to do in dnfdragora is that an on-demand fetch of changelog data for when people want to view changelogs. Since DNF doesn't support it, fetching it myself is the only option. This would let me be relatively cheap with that. And the way this is optimizing is not only really good, it extends well to *all* new metadata (like AppStream XML, which both Mageia and openSUSE append to the repodata). -- 真実はいつも一つ！/ Always, there's only one truth! _______________________________________________ Rpm-ecosystem mailing list Rpm-ecosystem@lists.rpm.org http://lists.rpm.org/mailman/listinfo/rpm-ecosystem
