Change Details

The main reason for the RFC tag is: the output of hg log -T {files} is stored in the changelog, despite it being a cache. So any fix to that cache changes commit hashes for impacted commits and all their descendants, which normally means most of the history (an exception being purely linear histories). So I wonder if this kind of change is acceptable, or whether some other approach is needed? (if this is ok, I should probably add a test with a criss-cross merge too to make sure the code is behaving as expected) Currently, the files list of merge commits repeats all the deletions (either actual deletions, or files that got renamed) that happened between base and p2 of the merge. If p2 is the main branch, the list can easily be much bigger than the change being merged. This impacts various areas of hg: - changelog becomes bigger than necessary - `hg log directory` lists many irrelevant commits - it possibly slows down adjustlinkrev, by forcing it to read more manifests, and that function can certainly be a bottleneck - the server side of pulls can spend a lot of time simply opening the filelogs for pointless files (the constant factors for opening even a tiny filelog is apparently pretty bad) So stop listing such files as described in the code. I don't have many concrete numbers to report, because recreating the files list of existing repositories is not easy: - debugupgradeformat and bundle/unbundle don't recreate the list - export/import tends to choke quickly applying patches or on description that contain diffs, - merge commits from the convert extension don't have the right files list for reasons orthogonal to the current commit - replaying the merge with hg update/hg merge/hg revert --all/hg commit can end up failing in hg revert - I wasn't sure that using debugsetparents + debugrebuilddirstate would really build the right thing I measured commit time before and after this change, in a case with no files filtered out, several files filtered out (no difference) and 5k files filtered out (+1% time). Recreating the 100 more recent merges in a private repo, the concatenated uncompressed files lists goes from 1.12MB to 0.52MB. Excluding 3 merges that are not representative, then the size goes from 570k to 15k. I converted part of mozilla-central, and observed file list shrinking quite a bit too (starting at the very first merge, 733641d9feaf, going from 550 files to 10 files), although they have relatively few merges, so they probably wouldn't care.

Currently, the files list of merge commits repeats all the deletions (either actual deletions, or files that got renamed) that happened between base and p2 of the merge. If p2 is the main branch, the list can easily be much bigger than the change being merged. This results in various problems worth improving: - changelog is bigger than necessary - `hg log directory` lists many unrelated merge commits, and `hg log -v -r commit` frequently fills multiple screens worth of files - it possibly slows down adjustlinkrev, by forcing it to read more manifests, and that function can certainly be a bottleneck - the server side of pulls can waste a lot of time simply opening the filelogs for pointless files (the constant factors for opening even a tiny filelog is apparently pretty bad) So stop listing such files as described in the code. Impacted merge commits and their descendants get a different hash than they would have without this. This doesn't seem problematic, except for convert. The previous commit helped with that in the hg->hg case (but if you do svn->hg twice from scratch, hashes can still change). The rest of the description is numbers. I don't have much to report, because recreating the files list of existing repositories is not easy: - debugupgradeformat and bundle/unbundle don't recreate the list - export/import tends to choke quickly applying patches or on description that contain diffs, - merge commits from the convert extension don't have the right files list for reasons orthogonal to the current commit - replaying the merge with hg update/hg merge/hg revert --all/hg commit can end up failing in hg revert - I wasn't sure that using debugsetparents + debugrebuilddirstate would really build the right thing I measured commit time before and after this change, in a case with no files filtered out, several files filtered out (no difference) and 5k files filtered out (+1% time). Recreating the 100 more recent merges in a private repo, the concatenated uncompressed files lists goes from 1.12MB to 0.52MB. Excluding 3 merges that are not representative, then the size goes from 570k to 15k. I converted part of mozilla-central, and observed file list shrinking quite a bit too, starting at the very first merge, 733641d9feaf, going from 550 files to 10 files (although they have relatively few merges, so they probably wouldn't care).

The main reason forCurrently, the RFC tag is: the output of hg log -T {files} isfiles list of merge commits repeats all the deletions stored in the changelog, despite it being a cache.(either actual deletions, So any fix toor files that got renamed) that happened cache changes commit hashes for impacted commits and allbetween base and p2 of the merge. If p2 is the main branch, their list descendants, which normally means most of the history (an exception being purely linear histories).can easily be much bigger than the change being merged. This results in various problems worth improving: So I wonder if this kind of change is acceptable, or whether some- changelog is bigger than necessary other approach is needed?- `hg log directory` lists many unrelated merge commits, and `hg log (if this is ok, I should probably add a test with a criss-cross merge -v -r commit` frequently fills multiple screens worth of files too to make sure the code is behaving as expected) Currently- it possibly slows down adjustlinkrev, the files list of merge commits repeats all the deletionsby forcing it to read more (either actual deletions manifests, or files that got renamed) that happenedand that function can certainly be a bottleneck between base and p2 of the merge. If p2 is the main branch,- the server side of pulls can waste a lot of time simply opening the list can easily be much bigger than the change being merged. This impacts various areas of hg: filelogs for pointless files (the constant factors for opening even - changelog becomes bigger than necessary - `hg log directory` lists many irrelevant commits - it possibly slows down adjustlinkrev, by forcing it to read mor a tiny filelog is apparently pretty bad) So stop listing such files as described in the code. Impacted merge manifests,commits and that function can certainly be a bottleneckeir descendants get a different hash than they would - the server side of pulls can spend a lot of time simply opening thehave without this. This doesn't seem problematic, except for filelogs for pointless files (the constant factors for opening evenconvert. The previous commit helped with that in the hg->hg case (but a tiny filelog is apparently pretty bad)if you do svn->hg twice from scratch, hashes can still change). So stop listing such files as described in the code. I don't have many concrete numbers to report, because recreating theThe rest of the description is numbers. I don't have much to report, because recreating the files list of existing repositories is not files list of existing repositories is not easy: - debugupgradeformat and bundle/unbundle don't recreate the list - export/import tends to choke quickly applying patches or on description that contain diffs, - merge commits from the convert extension don't have the right files list for reasons orthogonal to the current commit - replaying the merge with hg update/hg merge/hg revert --all/hg commit can end up failing in hg revert - I wasn't sure that using debugsetparents + debugrebuilddirstate would really build the right thing I measured commit time before and after this change, in a case with no files filtered out, several files filtered out (no difference) and 5k files filtered out (+1% time). Recreating the 100 more recent merges in a private repo, the concatenated uncompressed files lists goes from 1.12MB to 0.52MB. Excluding 3 merges that are not representative, then the size size goes from 570k to 15k. I converted part of mozilla-central, and observed file list shrinking quite a bit too (, starting at the very first merge, 733641d9feaf, going going from 550 files to 10 files), (although they have relatively few merges, few merges, so they probably wouldn't care).