Message ID | 20200529085038.26008-1-szeder.dev@gmail.com (mailing list archive) |
---|---|
Headers | show |
Series | An alternative modified path Bloom filters implementation | expand |
Hi Gábor, On Fri, 29 May 2020, SZEDER Gábor wrote: > Sigh... but better late than never, right? Well, incremental patches on top of what is _already_ in Git are _even_ better. > I experimented quite a bit with modified path Bloom filters a year and > more ago, and got quite far... but my disappointment in the > inadequacy of all double hashing schemes, the arrival of split > commit-graphs, and, well, life in general has put the whole thing on > the back burner, and I haven't touched it for a couple of releases. > > Now I finally managed to take a closer look at the current changed > paths Bloom filters implementation, and saw that it has some of the > same issues that I had stumbled upon and that it missed some > optimization opportunities. Unfortunately, fixing those issues and > performing those optimizations do require a thorough format change. Thank you for this background information. Your claim that there are opportunities to optimize, as well as your claim that it requires a thorough format change, strike me as best backed up with evidence by porting your optimizations on top of what is already in `master` and which has been reviewed _already_ (as opposed to your new patch series, which essentially threatens to throw away all the time people spent on reviewing Garima's patches). > So here is my proof of concept version, in all its incompleteness, > with the following benefits: > > - Better understanding of the problem it tries to optimize. > - Better understanding of the issues with many small Bloom filters. > - Better hashing scheme (though it should be better still). > - Orders of magnitude lower average false positive rate. > - Consequently, faster pathspec-limited revision walks. > - Faster processing of the tree-diff output and lower memory usage > while computing Bloom filters (from scratch...). > - Optional support for storing Bloom filters for all parents of > merge commits. > - Deduplicates Bloom filters. > - Supports multiple pathspecs right from the start. > - Supports some wildcards in pathspecs. > - Handles as many commits as the commit-graph format can. > - It has the right name :) The diff machinery and all its frontends > report "modified" paths with the letter 'M', not "changed". > - More cleanups, more bugfixes. > - Consistent output with and without modified path Bloom filters for > over 80k random paths in 16 repositories, even with submodules in > them. Well, at least on my machine, if nowhere else... It strikes me as an obvious idea to make all those improvements in an incremental manner. > Alas, the drawbacks are significant: > > - No tests whatsoever. > - Computes all modified path Bloom filters from scratch when > writing, no matter what. > - Doesn't work with split commit-graphs. > - Basically if anything works besides 'git commit-graph write > --reachable' it's a miracle. > - Not a single test. You said that already in the first bullet point. But yes, I get it, there are no tests. > - Many BUG()s, which should rather be graceful errors... though I > have to admit that at this point they are indeed bugs. > - Many TODOs, both in commit messages and code, some incomplete > commit messages, crappy subject lines, even missing signoffs. > - Some ridiculously long variable, function, macro and config > variable names. > - It's based on v2.25.0 (no technical reason, but that's the version > I used to run the baseline benchmarks the last time, which takes > days...) > - I'm pretty sure that there are more... > - Oh, did I mention that there are no tests? Ah, your patch series has no tests. Please do understand that it can be perceived as quite frustrating that an alternative is presented _this_ late, when already such a lot of effort has gone into not only iterating several times on the patch series but also into reviewing all of them. I don't really think that I want to spend the time to review this alternative (not that I am an expert in the space) because it would imply that I help invalidating the effort that went into the current implementation. All this is to say: I appreciate your efforts to improve the path-based Bloom filters, at the same time I wish that those efforts would happen in a collaborative manner instead of simply dismissing other people's work. Ciao, Johannes
On 5/29/2020 4:50 AM, SZEDER Gábor wrote: > Sigh... but better late than never, right? True. The best time to present these ideas would be while the series was under review, which was a reasonable amount of time. But there's no helping that, so what can we learn from your extensive work here? > I experimented quite a bit with modified path Bloom filters a year and > more ago, and got quite far... but my disappointment in the > inadequacy of all double hashing schemes, the arrival of split > commit-graphs, and, well, life in general has put the whole thing on > the back burner, and I haven't touched it for a couple of releases. > > Now I finally managed to take a closer look at the current changed > paths Bloom filters implementation, and saw that it has some of the > same issues that I had stumbled upon and that it missed some > optimization opportunities. Unfortunately, fixing those issues and > performing those optimizations do require a thorough format change. My initial reaction is "it's too late for a format change now" but that's not quite true. Even if we ship v2.27.0 with the existing format, the data is in optional chunks. It would not be too difficult to replace the storage with a different set of optional chunks and stop reading the old ones. The question remains: exactly _how much better_ is your implementation, and is it enough to justify the upheaval? The other thing that impacts this consideration is how complicated the format becomes. Bloom filters are already complicated, so how can we keep things simple when possible? > So here is my proof of concept version, in all its incompleteness, > with the following benefits: These are all very interesting items. Here is my initial categorization: These are bold claims that I hope you have end-to-end performance numbers to back them up. Depending on how much the benefit is, then we can consider a replacing. > - Better hashing scheme (though it should be better still). > - Orders of magnitude lower average false positive rate. > - Consequently, faster pathspec-limited revision walks. These bullets seem like they may apply to the existing implementation, since they seem unrelated to the format itself. Is it possible to contribute these ideas to the existing implementation first? > - Faster processing of the tree-diff output and lower memory usage > while computing Bloom filters (from scratch...). > - Supports multiple pathspecs right from the start. > - Supports some wildcards in pathspecs. > - More cleanups, more bugfixes. These items are confusing to me. I either don't know what they mean, or doubt the value of having them present. > - Better understanding of the problem it tries to optimize. > - Better understanding of the issues with many small Bloom filters. > - Deduplicates Bloom filters. > - It has the right name :) The diff machinery and all its frontends > report "modified" paths with the letter 'M', not "changed". > - Handles as many commits as the commit-graph format can> - Optional support for storing Bloom filters for all parents of > merge commits. (This last item in particular is one where I have already communicated why there is little value in storing filters for diffs against a second parent. Perhaps you have some concrete numbers on what happens in terms of storage-vs-performance when adding these diffs.) > Alas, the drawbacks are significant: > > - No tests whatsoever. > - Computes all modified path Bloom filters from scratch when > writing, no matter what. > - Doesn't work with split commit-graphs. > - Basically if anything works besides 'git commit-graph write > --reachable' it's a miracle. > - Not a single test. > - Many BUG()s, which should rather be graceful errors... though I > have to admit that at this point they are indeed bugs. > - Many TODOs, both in commit messages and code, some incomplete > commit messages, crappy subject lines, even missing signoffs. > - Some ridiculously long variable, function, macro and config > variable names. > - It's based on v2.25.0 (no technical reason, but that's the version > I used to run the baseline benchmarks the last time, which takes > days...) > - I'm pretty sure that there are more... > - Oh, did I mention that there are no tests? And now you get to the _real_ hard part of this feature. It is easy to create a prototype that implements the idea, but it's another step to make it a full contribution. You clearly know this, which I expect is the reason this was not submitted earlier. > > The first 14 patches are preparatory fixes and cleanups: > > 01/34 tree-walk.c: don't match submodule entries for 'submod/anything' > > This fix or something similar is necessary to have consistent output > with and without modified path Bloom filters for paths crossing > submodule boundary. > > 02/34 commit-graph: fix parsing the Chunk Lookup table > > The minimal (though not the best) fix for a bug which, I think, is as > old as the commit-graph. I don't know how to test this. > > 03/34 commit-graph-format.txt: all multi-byte numbers are in network byte order > 04/34 commit-slab: add a function to deep free entries on the slab > 05/34 diff.h: drop diff_tree_oid() & friends' return value > 06/34 commit-graph: clean up #includes > > A couple of minor cleanups. > > 07/34 commit-graph: simplify parse_commit_graph() #1 > 08/34 commit-graph: simplify parse_commit_graph() #2 > > These two would be the right, though not minimal fix for the parsing > bug above. > > 09/34 commit-graph: simplify write_commit_graph_file() #1 > 10/34 commit-graph: simplify write_commit_graph_file() #2 > 11/34 commit-graph: allocate the 'struct chunk_info' array dinamically > > I think these three cleanup patches are a better alternative of > 3be7efcafc (commit-graph: define and use MAX_NUM_CHUNKS, 2020-03-30), > because... > > 12/34 commit-graph: unify the signatures of all write_graph_chunk_*() functions > 13/34 commit-graph: simplify write_commit_graph_file() #3 > 14/34 commit-graph: check chunk sizes after writing These 14 patches are probably worth submitting on their own. I hope they still apply cleanly to a recent master. > ... they laid the ground work for this patch. > > 15/34 commit-graph-format.txt: document the modified path Bloom filter chunks > > This is the most important one, specifying and _justifying_ the new > chunk formats. > > Do grab a cup or pint of your favourite beverage and get comfy before > reading this one. You have been warned. > > 16/34 Add a generic and minimal Bloom filter implementation > 17/34 Import a streaming-capable Murmur3 hash function implementation > 18/34 commit-graph: write "empty" Modified Path Bloom Filter Index chunk > 19/34 commit-graph: add commit slab for modified path Bloom filters > 20/34 commit-graph: fill the Modified Path Bloom Filter Index chunk > > This shows a more efficient approach to process the tree-diff output > into Bloom filters. > > 21/34 commit-graph: load and use the Modified Path Bloom Filter Index chunk > 22/34 commit-graph: write the Modified Path Bloom Filters chunk > 23/34 commit-graph: load and use the Modified Path Bloom Filters chunk Ok, these implement the new filter format. > 24/34 commit-graph: check all leading directories in modified path Bloom filters > > This was a good lightbulb moment. It is essential to try to maintain > reasonable performance in repositories where the vast majority of > changes are concentrated to a single directory. This is really interesting! This could apply to the existing format, so when you compare formats this patch should either be in both or in neither. > 25/34 commit-graph: check embedded modified path Bloom filters with a mask > 26/34 commit-graph: deduplicate modified path Bloom filters > 27/34 commit-graph: load modified path Bloom filters for merge commits > 28/34 commit-graph: write Modified Path Bloom Filter Merge Index chunk > 29/34 commit-graph: extract init and free write_commit_graph_context > 30/34 commit-graph: move write_commit_graph_reachable below write_commit_graph > 31/34 t7007-show: make the first test compatible with the next patch > 32/34 PoC commit-graph: use revision walk machinery for '--reachable' > > Once upon a time I thought this was the greatest idea ever, but as > time goes by I get more and more concerned that this is a really dumb > idea, though don't yet know why. Ok, here are some extra patches that implement your extra features. The multi-parent filters are a fair thing to compare against the existing implementation, but we need to compare file size changes, extra computation time changes, and how much the end-to-end rev-list performance changes. Are the costs justified? > 33/34 commit-graph: write modified path Bloom filters in "history order" I think we've covered this in these two patches: 3d112755056 commit-graph: examine commits by generation number d21ee7d1110 commit-graph: examine changed-path objects in pack order > 34/34 commit-graph: use modified path Bloom filters with wildcards, if possible > > Finally a cherry on top. This is also something that could be interesting for the current implementation. Perhaps group it with patches 1-14 and 24, if the idea applies to the existing implementation. I think it would be great to apply whatever code cleanups and fixes that you've presented here to the existing implementation _first_, then we have the optimal way to compare how your proposed format improves over the existing implementation. I can't commit more time to this series today, but hopefully the discussion is just getting started. I'll try to poke around some of the patches next week. Thanks, -Stolee
On Fri, May 29, 2020 at 10:50:04AM +0200, SZEDER Gábor wrote: > Sigh... but better late than never, right? Yes, indeed. I think that there is a balance here: I'm thrilled that you are choosing to spend your time working on and improving the changed-path Bloom filter implementation. Of course, it couldn't have hurt to have these ideas earlier when the list was more focused on reviewing Garima's original patches. But, nothing is set in stone, and it seems like there are some re-usable ideas and clean-ups below. I'll respond to the patch descriptions below with a "cut-line" where I think we could extract and apply what's already there before putting the rest aside to be worked on more. > I experimented quite a bit with modified path Bloom filters a year and > more ago, and got quite far... but my disappointment in the > inadequacy of all double hashing schemes, the arrival of split > commit-graphs, and, well, life in general has put the whole thing on > the back burner, and I haven't touched it for a couple of releases. > > Now I finally managed to take a closer look at the current changed > paths Bloom filters implementation, and saw that it has some of the > same issues that I had stumbled upon and that it missed some > optimization opportunities. Unfortunately, fixing those issues and > performing those optimizations do require a thorough format change. > > So here is my proof of concept version, in all its incompleteness, > with the following benefits: > > - Better understanding of the problem it tries to optimize. > - Better understanding of the issues with many small Bloom filters. > - Better hashing scheme (though it should be better still). > - Orders of magnitude lower average false positive rate. > - Consequently, faster pathspec-limited revision walks. > - Faster processing of the tree-diff output and lower memory usage > while computing Bloom filters (from scratch...). > - Optional support for storing Bloom filters for all parents of > merge commits. > - Deduplicates Bloom filters. > - Supports multiple pathspecs right from the start. > - Supports some wildcards in pathspecs. > - Handles as many commits as the commit-graph format can. > - It has the right name :) The diff machinery and all its frontends > report "modified" paths with the letter 'M', not "changed". > - More cleanups, more bugfixes. > - Consistent output with and without modified path Bloom filters for > over 80k random paths in 16 repositories, even with submodules in > them. Well, at least on my machine, if nowhere else... > > Alas, the drawbacks are significant: > > - No tests whatsoever. Yes, obviously this needs to be addressed, but I certainly don't fault you for posting what you have (thanks for adding the 'PoC' prefix to indicate it as such). > - Computes all modified path Bloom filters from scratch when > writing, no matter what. > - Doesn't work with split commit-graphs. I originally thought that this would be a non-starter, but it may not be. GitHub doesn't write Bloom filters at all (yet), but one consideration of ours is to only generate Bloom filters when we roll-up all of the split graphs. Our invocation during this bigger repository maintenance is something like: $ git commit-graph write --reachable --split=replace But if it turns out to be expensive to run 'git commit-graph write --split=no-merge --stdin-commits --write --changed-paths' on every push, we might just run the above with '--changed-paths'. I can imagine other schemes where it would be desirable to have your Bloom filter implementation over split graphs, in which case this should be addressed. But, I don't have a problem with a version that only works for non-split graphs first, so long as there isn't a compatibility boundary between that and a version that does work for split graphs. > - Basically if anything works besides 'git commit-graph write > --reachable' it's a miracle. > - Not a single test. > - Many BUG()s, which should rather be graceful errors... though I > have to admit that at this point they are indeed bugs. > - Many TODOs, both in commit messages and code, some incomplete > commit messages, crappy subject lines, even missing signoffs. > - Some ridiculously long variable, function, macro and config > variable names. > - It's based on v2.25.0 (no technical reason, but that's the version > I used to run the baseline benchmarks the last time, which takes > days...) > - I'm pretty sure that there are more... > - Oh, did I mention that there are no tests? ...I think so ;-). > > The first 14 patches are preparatory fixes and cleanups: > > 01/34 tree-walk.c: don't match submodule entries for 'submod/anything' > > This fix or something similar is necessary to have consistent output > with and without modified path Bloom filters for paths crossing > submodule boundary. > > 02/34 commit-graph: fix parsing the Chunk Lookup table > > The minimal (though not the best) fix for a bug which, I think, is as > old as the commit-graph. I don't know how to test this. > > 03/34 commit-graph-format.txt: all multi-byte numbers are in network byte order > 04/34 commit-slab: add a function to deep free entries on the slab > 05/34 diff.h: drop diff_tree_oid() & friends' return value > 06/34 commit-graph: clean up #includes > > A couple of minor cleanups. > > 07/34 commit-graph: simplify parse_commit_graph() #1 > 08/34 commit-graph: simplify parse_commit_graph() #2 > > These two would be the right, though not minimal fix for the parsing > bug above. > > 09/34 commit-graph: simplify write_commit_graph_file() #1 > 10/34 commit-graph: simplify write_commit_graph_file() #2 > 11/34 commit-graph: allocate the 'struct chunk_info' array dinamically > > I think these three cleanup patches are a better alternative of > 3be7efcafc (commit-graph: define and use MAX_NUM_CHUNKS, 2020-03-30), > because... > > 12/34 commit-graph: unify the signatures of all write_graph_chunk_*() functions > 13/34 commit-graph: simplify write_commit_graph_file() #3 > 14/34 commit-graph: check chunk sizes after writing I think you're right to draw the "laying the ground work" line here. Could these first fourteen patches be applied cleanly to master? They all look mostly like improvements to me, especially the second patch. > ... they laid the ground work for this patch. > > 15/34 commit-graph-format.txt: document the modified path Bloom filter chunks > > This is the most important one, specifying and _justifying_ the new > chunk formats. > > Do grab a cup or pint of your favourite beverage and get comfy before > reading this one. You have been warned. > > 16/34 Add a generic and minimal Bloom filter implementation > 17/34 Import a streaming-capable Murmur3 hash function implementation > 18/34 commit-graph: write "empty" Modified Path Bloom Filter Index chunk > 19/34 commit-graph: add commit slab for modified path Bloom filters > 20/34 commit-graph: fill the Modified Path Bloom Filter Index chunk > > This shows a more efficient approach to process the tree-diff output > into Bloom filters. > > 21/34 commit-graph: load and use the Modified Path Bloom Filter Index chunk > 22/34 commit-graph: write the Modified Path Bloom Filters chunk > 23/34 commit-graph: load and use the Modified Path Bloom Filters chunk > 24/34 commit-graph: check all leading directories in modified path Bloom filters > > This was a good lightbulb moment. It is essential to try to maintain > reasonable performance in repositories where the vast majority of > changes are concentrated to a single directory. > > 25/34 commit-graph: check embedded modified path Bloom filters with a mask > 26/34 commit-graph: deduplicate modified path Bloom filters > 27/34 commit-graph: load modified path Bloom filters for merge commits > 28/34 commit-graph: write Modified Path Bloom Filter Merge Index chunk > 29/34 commit-graph: extract init and free write_commit_graph_context > 30/34 commit-graph: move write_commit_graph_reachable below write_commit_graph > 31/34 t7007-show: make the first test compatible with the next patch > 32/34 PoC commit-graph: use revision walk machinery for '--reachable' > > Once upon a time I thought this was the greatest idea ever, but as > time goes by I get more and more concerned that this is a really dumb > idea, though don't yet know why. > > 33/34 commit-graph: write modified path Bloom filters in "history order" > 34/34 commit-graph: use modified path Bloom filters with wildcards, if possible > > Finally a cherry on top. The next twenty patches look like they need some more work. At a high-level, I think we need to do the following things, in roughly the following order: - Prove that this is a strict performance improvement on the existing Bloom filter implementation. - Sketch out a way that it would work for cases where '--split' is provided, and/or '--reachable' isn't provided. - Clean up the existing patches, adding tests which exercise the new functionality, and prove that there aren't regressions against the existing Bloom filter implementation. - Review and repeat. I guess I'm left wondering whether or not this is something that you want to continue, given that I outlined a pretty lengthy path for getting this in a suitable shape to keep going. Are you planning on keeping working on this? Thanks, Taylor
Taylor Blau <me@ttaylorr.com> writes: > On Fri, May 29, 2020 at 10:50:04AM +0200, SZEDER Gábor wrote: >> Sigh... but better late than never, right? > > Yes, indeed. I think that there is a balance here: I'm thrilled that you > are choosing to spend your time working on and improving the > changed-path Bloom filter implementation. > > Of course, it couldn't have hurt to have these ideas earlier when the > list was more focused on reviewing Garima's original patches. But, > nothing is set in stone, and it seems like there are some re-usable > ideas and clean-ups below. Yes, I had the same impression as you did, unlike some folks who sounded as if they felt offended seeing a comment that sabotages existing work. It could have been presented in a more useful ways (i.e. instead of risking to appear suggesting total replacement, which I do not think was the intention, massaged to build on top of what is already there as improvements), though. > I think you're right to draw the "laying the ground work" line here. > Could these first fourteen patches be applied cleanly to master? They > all look mostly like improvements to me, especially the second patch. Again, I concur. Thanks.