| Message ID | a873ed828ccae426214cc8f87610df97ff9a269e.1729055871.git.ps@pks.im (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | ref-filter: format iteratively with lexicographic refname sorting | expand |
On Wed, Oct 16, 2024 at 08:00:30AM +0200, Patrick Steinhardt wrote: > But there is one exception here where we _can_ get away with sorting > refs while streaming: ref backends sort references returned by their > iterators in lexicographic order. So if the following conditions are all > true we can do iterative streaming: > > - The caller uses at most a single name pattern. Otherwise we'd have > to sort results from multiple invocations of the iterator. > > - There must be at most a single sorting specification, as otherwise > we're not using plain lexicographic ordering. > > - The sorting specification must use the "refname". > > - The sorting specification must not be using any flags, like > case-insensitive sorting. Perhaps a niche case, but what about ancient packed-refs files that were written before the 'sorted' capability was introduced? Thanks, Taylor
On Wed, Oct 16, 2024 at 06:11:47PM -0400, Taylor Blau wrote: > On Wed, Oct 16, 2024 at 08:00:30AM +0200, Patrick Steinhardt wrote: > > But there is one exception here where we _can_ get away with sorting > > refs while streaming: ref backends sort references returned by their > > iterators in lexicographic order. So if the following conditions are all > > true we can do iterative streaming: > > > > - The caller uses at most a single name pattern. Otherwise we'd have > > to sort results from multiple invocations of the iterator. > > > > - There must be at most a single sorting specification, as otherwise > > we're not using plain lexicographic ordering. > > > > - The sorting specification must use the "refname". > > > > - The sorting specification must not be using any flags, like > > case-insensitive sorting. > > Perhaps a niche case, but what about ancient packed-refs files that were > written before the 'sorted' capability was introduced? We should be OK there. In that case we actually read in and sort the packed-refs entries ourselves. We have to, since we do an in-order merge with the loose refs while iterating. I do think this optimization is worth doing, and not a problem with our current backends. The biggest worries would be: 1. Some new ref backend that doesn't return sorted results. I find this unlikely, and anyway it's easily caught by having coverage in the test suite (which I assume we already have, but I didn't look). 2. Some new flag combination that requires disabling the optimization, and which must be dealt with in the code. This seems unlikely to me but not impossible. I think enabling the optimization is worth it, though. > > - The caller uses at most a single name pattern. Otherwise we'd have > > to sort results from multiple invocations of the iterator. I think this part is erring on the cautious side, as we can often collapse these into a single iteration due to the ref-prefix work. It may be OK to keep using the slower code here if multiple patterns aren't commonly used, but I'd suspect that: git for-each-ref --sort=refname refs/heads refs/tags could benefit. -Peff
On Wed, Oct 16, 2024 at 10:48:28PM -0400, Jeff King wrote: > On Wed, Oct 16, 2024 at 06:11:47PM -0400, Taylor Blau wrote: > > > On Wed, Oct 16, 2024 at 08:00:30AM +0200, Patrick Steinhardt wrote: > > > But there is one exception here where we _can_ get away with sorting > > > refs while streaming: ref backends sort references returned by their > > > iterators in lexicographic order. So if the following conditions are all > > > true we can do iterative streaming: > > > > > > - The caller uses at most a single name pattern. Otherwise we'd have > > > to sort results from multiple invocations of the iterator. > > > > > > - There must be at most a single sorting specification, as otherwise > > > we're not using plain lexicographic ordering. > > > > > > - The sorting specification must use the "refname". > > > > > > - The sorting specification must not be using any flags, like > > > case-insensitive sorting. > > > > Perhaps a niche case, but what about ancient packed-refs files that were > > written before the 'sorted' capability was introduced? > > We should be OK there. In that case we actually read in and sort the > packed-refs entries ourselves. We have to, since we do an in-order merge > with the loose refs while iterating. > > I do think this optimization is worth doing, and not a problem with our > current backends. The biggest worries would be: > > 1. Some new ref backend that doesn't return sorted results. I find > this unlikely, and anyway it's easily caught by having coverage in > the test suite (which I assume we already have, but I didn't look). My assumption is that a ref iterator that _isn't_ sorted would lead to undesirable behaviour. I'd be surprised if git-show-ref(1) started to show refs in a random order. So we have essentially baked the requirement that ref iterators return refs in lexicographic order into our codebase already. > 2. Some new flag combination that requires disabling the optimization, > and which must be dealt with in the code. This seems unlikely to me > but not impossible. I think enabling the optimization is worth it, > though. And this part is an issue with or without my patch. The logic we have in the ref-filter API is quite fragile, and everybody who wants to add some new flags must remember to update `can_do_iterative_format()`. I'm not really a huge fan of that, but on the other hand the subsystem does not change all that frequently anyway. > > > to sort results from multiple invocations of the iterator. > > I think this part is erring on the cautious side, as we can often > collapse these into a single iteration due to the ref-prefix work. It > may be OK to keep using the slower code here if multiple patterns aren't > commonly used, but I'd suspect that: > > git for-each-ref --sort=refname refs/heads refs/tags > > could benefit. Mh. So we do end up using `refs_for_each_fullref_in_prefixes()`, which may or may not end up collapsing the prefixes. We do sort and dedup the prefixes via `find_longest_prefixes()`, so we don't have to worry about e.g. `git for-each-ref refs/tags refs/heads refs/tags`. So... it should be fine to also use this with multiple patterns? None of our tests fail, either, which reassures me a bit. I'll send a v2 that loosens this requirement. Thanks for your input! Patrick
diff --git a/ref-filter.c b/ref-filter.c index dd195007ce1..e075ca21d8e 100644 --- a/ref-filter.c +++ b/ref-filter.c @@ -3244,10 +3244,40 @@ int filter_refs(struct ref_array *array, struct ref_filter *filter, unsigned int return ret; } +struct ref_sorting { + struct ref_sorting *next; + int atom; /* index into used_atom array (internal) */ + enum ref_sorting_order sort_flags; +}; + static inline int can_do_iterative_format(struct ref_filter *filter, struct ref_sorting *sorting, struct ref_format *format) { + /* + * Reference backends sort patterns lexicographically by refname, so if + * the sorting options ask for exactly that we may be able to do + * iterative formatting. + */ + if (sorting) { + size_t n = 0; + + /* + * There must be a single sorting filter that uses + * lexicographic sorting of the refname. + */ + if (sorting->next || + sorting->sort_flags || + used_atom[sorting->atom].atom_type != ATOM_REFNAME) + return 0; + + /* And there must be at most a single name pattern. */ + while (filter->name_patterns && filter->name_patterns[n] && n < 2) + n++; + if (n > 1) + return 0; + } + /* * Filtering & formatting results within a single ref iteration * callback is not compatible with options that require @@ -3258,7 +3288,6 @@ static inline int can_do_iterative_format(struct ref_filter *filter, */ return !(filter->reachable_from || filter->unreachable_from || - sorting || format->bases.nr || format->is_base_tips.nr); } @@ -3316,12 +3345,6 @@ static int memcasecmp(const void *vs1, const void *vs2, size_t n) return 0; } -struct ref_sorting { - struct ref_sorting *next; - int atom; /* index into used_atom array (internal) */ - enum ref_sorting_order sort_flags; -}; - static int cmp_ref_sorting(struct ref_sorting *s, struct ref_array_item *a, struct ref_array_item *b) { struct atom_value *va, *vb;
In bd98f9774e (ref-filter.c: filter & format refs in the same callback, 2023-11-14), we have introduced logic into the ref-filter subsystem that determines whether or not we can output references iteratively instead of first collecting all references, post-processing them and printing them once done. This has the advantage that we don't have to store all refs in memory and, when used with e.g. `--count=1`, that we don't have to read all refs in the first place. One restriction we have in place for that is that caller must not ask for sorted refs, because there is no way to sort the refs without first reading them all into an array. So the benefits can only be reaped when explicitly asking for output not to be sorted. But there is one exception here where we _can_ get away with sorting refs while streaming: ref backends sort references returned by their iterators in lexicographic order. So if the following conditions are all true we can do iterative streaming: - The caller uses at most a single name pattern. Otherwise we'd have to sort results from multiple invocations of the iterator. - There must be at most a single sorting specification, as otherwise we're not using plain lexicographic ordering. - The sorting specification must use the "refname". - The sorting specification must not be using any flags, like case-insensitive sorting. Now the resulting logic does feel quite fragile overall, which makes me a bit uneasy. But after thinking about this for a while I couldn't find any obvious gaps in my reasoning. Furthermore, given that lexicographic sorting order is the default in git-for-each-ref(1), this is likely to benefit a whole lot of usecases out there. The following benchmark executes git-for-each-ref(1) in a crafted repo with 1 million references: Benchmark 1: git for-each-ref (revision = HEAD~) Time (mean ± σ): 6.756 s ± 0.014 s [User: 3.004 s, System: 3.541 s] Range (min … max): 6.738 s … 6.784 s 10 runs Benchmark 2: git for-each-ref (revision = HEAD) Time (mean ± σ): 6.479 s ± 0.017 s [User: 2.858 s, System: 3.422 s] Range (min … max): 6.450 s … 6.519 s 10 runs Summary git for-each-ref (revision = HEAD) 1.04 ± 0.00 times faster than git for-each-ref (revision = HEAD~) The change results in a slight performance improvement, but nothing that would really stand out. Something that cannot be seen in the benchmark though is peak memory usage, which went from 404.5MB to 68.96kB. A more interesting benchmark is printing a single referenence with `--count=1`: Benchmark 1: git for-each-ref --count=1 (revision = HEAD~) Time (mean ± σ): 6.655 s ± 0.018 s [User: 2.865 s, System: 3.576 s] Range (min … max): 6.630 s … 6.680 s 10 runs Benchmark 2: git for-each-ref --count=1 (revision = HEAD) Time (mean ± σ): 8.6 ms ± 1.3 ms [User: 2.3 ms, System: 6.1 ms] Range (min … max): 6.7 ms … 14.4 ms 266 runs Summary git git for-each-ref --count=1 (revision = HEAD) 770.58 ± 116.19 times faster than git for-each-ref --count=1 (revision = HEAD~) Whereas we scaled with the number of references before, we now print the first reference and exit immediately, which provides a massive win. Signed-off-by: Patrick Steinhardt <ps@pks.im> --- I'm honestly not quite sure whether I think that this change is fine, or whether it is getting too fragile. I decided to send the patch anyway so that we can discuss on the mailing list, mostly because I think that the results speak for themselves. Ultimately, this very much feels like a tradeoff to me. Thanks! Patrick ref-filter.c | 37 ++++++++++++++++++++++++++++++------- 1 file changed, 30 insertions(+), 7 deletions(-)