# Intro Last year, John Cai sent 2 versions of a patch series to implement `git repack --filter=<filter-spec>` and later I sent 4 versions of a patch series trying to do it a bit differently: - https://lore.kernel.org/git/pull.1206.git.git.1643248180.gitgitgadget@xxxxxxxxx/ - https://lore.kernel.org/git/20221012135114.294680-1-christian.couder@xxxxxxxxx/ In these patch series, the `--filter=<filter-spec>` removed the filtered out objects altogether which was considered very dangerous even though we implemented different safety checks in some of the latter series. In some discussions, it was mentioned that such a feature, or a similar feature in `git gc`, or in a new standalone command (perhaps called `git prune-filtered`), should put the filtered out objects into a new packfile instead of deleting them. Recently there were internal discussions at GitLab about either moving blobs from inactive repos onto cheaper storage, or moving large blobs onto cheaper storage. This lead us to rethink at repacking using a filter, but moving the filtered out objects into a separate packfile instead of deleting them. So here is a new patch series doing that while implementing the `--filter=<filter-spec>` option in `git repack`. # Use cases for the new feature This could be useful for example for the following purposes: 1) As a way for servers to save storage costs by for example moving large blobs, or all the blobs, or all the blobs in inactive repos, to separate storage (while still making them accessible using for example the alternates mechanism). 2) As a way to use partial clone on a Git server to offload large blobs to, for example, an http server, while using multiple promisor remotes (to be able to access everything) on the client side. (In this case the packfile that contains the filtered out object can be manualy removed after checking that all the objects it contains are available through the promisor remote.) 3) As a way for clients to reclaim some space when they cloned with a filter to save disk space but then fetched a lot of unwanted objects (for example when checking out old branches) and now want to remove these unwanted objects. (In this case they can first move the packfile that contains filtered out objects to a separate directory or storage, then check that everything works well, and then manually remove the packfile after some time.) As the features and the code are quite different from those in the previous series, I decided to start a new series instead of continuing a previous one. Also since version 2 of this new series, commit messages, don't mention uses cases like 2) or 3) above, as people have different opinions on how it should be done. How it should be done could depend a lot on the way promisor remotes are used, the software and hardware setups used, etc, so it seems more difficult to "sell" this series by talking about such use cases. As use case 1) seems simpler and more appealing, it makes more sense to only talk about it in the commit messages. # Changes since version 7 Thanks to Junio who reviewed or commented on nearly all the versions, and to Taylor who reviewed or commented on version 1, 3, 4, 5 and 6! Thanks also to Robert Coup who participated in the discussions related to version 2 and Peff who participated in the discussions related to version 4. There are only the following changes since version 7: - This series has been rebased on top of 493f462273 (The thirteenth batch, 2023-09-29) to avoid possible conflicts with other series that could potentially conflict with this one. - Patch 2/9 (t/helper: add 'find-pack' test-tool) has been reworked to use the "t0081" test script number instead of "t0080" as the later is used by js/doc-unit-tests. I asked in: https://lore.kernel.org/git/CAP8UFD2YbYH5aZEG5NX8HLe9VeEQ+NhBfiZ9Mhy3UXTUrab3ug@xxxxxxxxxxxxxx/ if someone thought another number was better, but got no answer. I checked that CI tests passes in: https://github.com/chriscool/git/actions/runs/6382343338 All jobs seem to have succeeded. # Commit overview (No changes in any of the patches compared to version 7, except on patch 2/9.) * 1/9 pack-objects: allow `--filter` without `--stdout` To be able to later repack with a filter we need `git pack-objects` to write packfiles when it's filtering instead of just writing the pack without the filtered out objects to stdout. * 2/9 t/helper: add 'find-pack' test-tool For testing `git repack --filter=...` that we are going to implement, it's useful to have a test helper that can tell which packfiles contain a specific object. The only change compared to v7 is the change in test script number. * 3/9 repack: refactor finishing pack-objects command This is a small refactoring creating a new useful function, so that `git repack --filter=...` will be able to reuse it. * 4/9 repack: refactor finding pack prefix This is another small refactoring creating a small function that will be reused in the next patch. * 5/9 pack-bitmap-write: rebuild using new bitmap when remapping It fixes an issue when bitmaps are rebuilt that was revealed by this series, and caused a CI test to fail. * 6/9 repack: add `--filter=<filter-spec>` option This actually adds the `--filter=<filter-spec>` option. It uses one `git pack-objects` process with the `--filter` option. And then another `git pack-objects` process with the `--stdin-packs` option. * 7/9 gc: add `gc.repackFilter` config option This is a gc config option so that `git gc` can also repack using a filter and put the filtered out objects into a separate packfile. * 8/9 repack: implement `--filter-to` for storing filtered out objects For some use cases, it's interesting to create the packfile that contains the filtered out objects into a separate location. This is similar to the `--expire-to` option for cruft packfiles. * 9/9 gc: add `gc.repackFilterTo` config option This allows specifying the location of the packfile that contains the filtered out objects when using `gc.repackFilter`. # Range-diff since v7 1: eec0c09731 = 1: b23d216277 pack-objects: allow `--filter` without `--stdout` 2: 19c8b8a4b9 ! 2: 27e70ccf39 t/helper: add 'find-pack' test-tool @@ t/helper/test-tool.h: int cmd__dump_reftable(int argc, const char **argv); int cmd__genrandom(int argc, const char **argv); int cmd__genzeros(int argc, const char **argv); - ## t/t0080-find-pack.sh (new) ## + ## t/t0081-find-pack.sh (new) ## @@ +#!/bin/sh + 3: aaaf40bd5d = 3: 7e692c4cfd repack: refactor finishing pack-objects command 4: 1eb6bc3f7e = 4: 227159ed4e repack: refactor finding pack prefix 5: b9159e1803 = 5: 79786eb5e1 pack-bitmap-write: rebuild using new bitmap when remapping 6: f2f5bb54d3 = 6: 205d33850e repack: add `--filter=<filter-spec>` option 7: 7ea0307628 = 7: 16b1621169 gc: add `gc.repackFilter` config option 8: 698647815b = 8: 92a5ff7cc7 repack: implement `--filter-to` for storing filtered out objects 9: 57b2ba444c = 9: 5bfd918c90 gc: add `gc.repackFilterTo` config option Christian Couder (9): pack-objects: allow `--filter` without `--stdout` t/helper: add 'find-pack' test-tool repack: refactor finishing pack-objects command repack: refactor finding pack prefix pack-bitmap-write: rebuild using new bitmap when remapping repack: add `--filter=<filter-spec>` option gc: add `gc.repackFilter` config option repack: implement `--filter-to` for storing filtered out objects gc: add `gc.repackFilterTo` config option Documentation/config/gc.txt | 16 ++ Documentation/git-pack-objects.txt | 4 +- Documentation/git-repack.txt | 23 +++ Makefile | 1 + builtin/gc.c | 10 ++ builtin/pack-objects.c | 8 +- builtin/repack.c | 164 ++++++++++++++------ pack-bitmap-write.c | 6 +- t/helper/test-find-pack.c | 50 +++++++ t/helper/test-tool.c | 1 + t/helper/test-tool.h | 1 + t/t0081-find-pack.sh | 82 ++++++++++ t/t5317-pack-objects-filter-objects.sh | 8 + t/t6500-gc.sh | 24 +++ t/t7700-repack.sh | 197 +++++++++++++++++++++++++ 15 files changed, 544 insertions(+), 51 deletions(-) create mode 100644 t/helper/test-find-pack.c create mode 100755 t/t0081-find-pack.sh -- 2.42.0.305.g5bfd918c90
