RE: [PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
> -Original Message- > From: Ben Peart [mailto:peart...@gmail.com] > Sent: Thursday, May 18, 2017 12:58 PM > To: Junio C Hamano <gits...@pobox.com>; David Turner > <david.tur...@twosigma.com> > Cc: 'Christian Couder' <christian.cou...@gmail.com>; Johannes Schindelin > <johannes.schinde...@gmx.de>; git <git@vger.kernel.org>; Nguyễn Thái Ngọc > Duy <pclo...@gmail.com>; Ben Peart <benpe...@microsoft.com> > Subject: Re: [PATCH 0/1] Preserve the untracked cache across checkout, reset > -- > hard, etc > > > > On 5/9/2017 8:51 AM, Ben Peart wrote: > > > > On 5/9/2017 1:02 AM, Junio C Hamano wrote: > >> David Turner <david.tur...@twosigma.com> writes: > >> > >>> Can you actually keep the email address as my Twopensource one? I > >>> want to make sure that Twitter, my employer at the time, gets credit > >>> for this work (just as I want to make sure that my current employer, > >>> Two Sigma, gets credit for my current work). > >>> > >>> Please feel free to add Signed-off-by: David Turner > >>> <dtur...@twosigma.com> in case that makes tracking easier. > >>> > >>> Thanks. > >>> > >>> WRT the actual patch, I want to note that past me did not do a great > >>> job here. The tests do not correctly check that the post-checkout > >>> untracked cache is still valid after a checkout. > >>> For example, let's say that previously, the directory foo was > >>> entirely untracked (but it contained a file bar), but after the > >>> checkout, there is a file foo/baz. Does the untracked cache need to > >>> get updated? > >>> > >>> Unfortunately, the untracked cache is very unlikely to make it to > >>> the top of my priority list any time soon, so I won't be able to > >>> correct this test (and, if necessary, correct the code). But I > >>> would strongly suggest that the test be improved before this code is > >>> merged. > >>> > >>> Thanks for CCing me. > >> I will try to find time to tweak what was sent to the list here to > >> reflect your affiliations better, but marked with DONTMERGE waiting > >> for the necessary updates you mentioned above, so that this change is > >> not forgotten. It may turn out to be that copying from src to dst > >> like the patch does is all that is needed, or the cache may need > >> further invalidation when the copying happens, and I haven't got a > >> good feeling that anybody who are familiar with the codepath vetted > >> the correctness from seeing the discussion from sidelines (yet). > >> > >> Thanks. > > > > I've been looking into similar issues with the cache associated with > > using a file system monitor (aka Watchman) > > (https://github.com/git-for-windows/git/compare/master...benpeart:fsmo > > nitor) to speed updating the index to reflect changes in the working > > directory. > > > > I can take a look and see if this patch results in valid results and > > reply to the thread or resubmit as needed. > > > > Ben > > TLDR: the patch looks good from my perspective but I'd like the experts to > weigh > in as well. Thanks for looking into this. I'm glad to learn that I got it right the first time, although I still wish I had been more assiduous about testing back then. > After digging into the untracked cache code and thinking about whether it is > reasonable to copy the cache from the old index to the new index in > unpack_trees() I believe the answer is "yes." I'm not the expert in this > code so I'll > outline my reasoning here and hopefully the real experts can review it and > see if > I've missed something. > > The interesting part of the untracked cache for this discussion is the list of > untracked_cache_dir structures. Because each directory cache entry contains > stat_data (esp ctime and mtime) for that directory - the existing logic will > detect > if that directory has had any changes made in it since the cache entry was > saved. > It doesn't really care when, why, or how the change was made, just if one has > happened. > > I then tried to think of ways that this logic could be broken (like David's > example > above) but was unsuccessful in coming up with any. This makes sense because > the untracked cache obviously has to correctly detect _any_ change so really > doesn't care whether it's cached state was initially saved before or after a > call to > unpack_trees(). It looks like un
Re: [PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
On 5/9/2017 8:51 AM, Ben Peart wrote: On 5/9/2017 1:02 AM, Junio C Hamano wrote: David Turnerwrites: Can you actually keep the email address as my Twopensource one? I want to make sure that Twitter, my employer at the time, gets credit for this work (just as I want to make sure that my current employer, Two Sigma, gets credit for my current work). Please feel free to add Signed-off-by: David Turner in case that makes tracking easier. Thanks. WRT the actual patch, I want to note that past me did not do a great job here. The tests do not correctly check that the post-checkout untracked cache is still valid after a checkout. For example, let's say that previously, the directory foo was entirely untracked (but it contained a file bar), but after the checkout, there is a file foo/baz. Does the untracked cache need to get updated? Unfortunately, the untracked cache is very unlikely to make it to the top of my priority list any time soon, so I won't be able to correct this test (and, if necessary, correct the code). But I would strongly suggest that the test be improved before this code is merged. Thanks for CCing me. I will try to find time to tweak what was sent to the list here to reflect your affiliations better, but marked with DONTMERGE waiting for the necessary updates you mentioned above, so that this change is not forgotten. It may turn out to be that copying from src to dst like the patch does is all that is needed, or the cache may need further invalidation when the copying happens, and I haven't got a good feeling that anybody who are familiar with the codepath vetted the correctness from seeing the discussion from sidelines (yet). Thanks. I've been looking into similar issues with the cache associated with using a file system monitor (aka Watchman) (https://github.com/git-for-windows/git/compare/master...benpeart:fsmonitor) to speed updating the index to reflect changes in the working directory. I can take a look and see if this patch results in valid results and reply to the thread or resubmit as needed. Ben TLDR: the patch looks good from my perspective but I'd like the experts to weigh in as well. After digging into the untracked cache code and thinking about whether it is reasonable to copy the cache from the old index to the new index in unpack_trees() I believe the answer is "yes." I'm not the expert in this code so I'll outline my reasoning here and hopefully the real experts can review it and see if I've missed something. The interesting part of the untracked cache for this discussion is the list of untracked_cache_dir structures. Because each directory cache entry contains stat_data (esp ctime and mtime) for that directory - the existing logic will detect if that directory has had any changes made in it since the cache entry was saved. It doesn't really care when, why, or how the change was made, just if one has happened. I then tried to think of ways that this logic could be broken (like David's example above) but was unsuccessful in coming up with any. This makes sense because the untracked cache obviously has to correctly detect _any_ change so really doesn't care whether it's cached state was initially saved before or after a call to unpack_trees(). Even scenarios of creating files in sub-directories of sub-directories works because eventually, either is a directory or file is created in a cached directory entry which will change the mtime of that directory and invalidate that part of the cache. Ultimately, it is this behavior of saving the mtime of each cached directory that makes this all work as each entry can be validated/invalidated separately from all the rest and independently from the index from which they came. Once I did the code examination and thinking exercise, I wanted to test it out and see if the theory held up. I started out with some manual testing (esp of the scenario David mentioned) and then wrote a couple of additional tests all of which passed. I then ran all existing git tests with the patch applied and they all passed. This only really tells us that it didn't break anything because untracked cache is turned off by default but it does show us that it still passes the untracked cache specific test cases (as they obviously turn it on). I then modified the test_create_repo() function in test-lib-functions.sh to turn on the untracked cache feature after creating the test repo and ran all the tests again twice - the first time without the patch and again with the patch). This run is more interesting because it is testing that having the untracked cache turned (with and without the patch) on doesn't break anything. There were two test scripts that had failures: t7063-status-untracked-cache.sh failed the test "not ok 1 - core.untrackedCache is unset" This is actually a positive result because it is showing that I successfully turned
Re: [PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
On 5/9/2017 1:02 AM, Junio C Hamano wrote: David Turnerwrites: Can you actually keep the email address as my Twopensource one? I want to make sure that Twitter, my employer at the time, gets credit for this work (just as I want to make sure that my current employer, Two Sigma, gets credit for my current work). Please feel free to add Signed-off-by: David Turner in case that makes tracking easier. Thanks. WRT the actual patch, I want to note that past me did not do a great job here. The tests do not correctly check that the post-checkout untracked cache is still valid after a checkout. For example, let's say that previously, the directory foo was entirely untracked (but it contained a file bar), but after the checkout, there is a file foo/baz. Does the untracked cache need to get updated? Unfortunately, the untracked cache is very unlikely to make it to the top of my priority list any time soon, so I won't be able to correct this test (and, if necessary, correct the code). But I would strongly suggest that the test be improved before this code is merged. Thanks for CCing me. I will try to find time to tweak what was sent to the list here to reflect your affiliations better, but marked with DONTMERGE waiting for the necessary updates you mentioned above, so that this change is not forgotten. It may turn out to be that copying from src to dst like the patch does is all that is needed, or the cache may need further invalidation when the copying happens, and I haven't got a good feeling that anybody who are familiar with the codepath vetted the correctness from seeing the discussion from sidelines (yet). Thanks. I've been looking into similar issues with the cache associated with using a file system monitor (aka Watchman) (https://github.com/git-for-windows/git/compare/master...benpeart:fsmonitor) to speed updating the index to reflect changes in the working directory. I can take a look and see if this patch results in valid results and reply to the thread or resubmit as needed. Ben
Re: [PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
David Turnerwrites: > Can you actually keep the email address as my Twopensource one? I want to > make sure that Twitter, my employer at the time, gets credit for this work > (just as I want to make sure that my current employer, Two Sigma, gets credit > for my current work). > > Please feel free to add Signed-off-by: David Turner in > case that makes tracking easier. > > Thanks. > > WRT the actual patch, I want to note that past me did not do a > great job here. The tests do not correctly check that the > post-checkout untracked cache is still valid after a checkout. > For example, let's say that previously, the directory foo was > entirely untracked (but it contained a file bar), but after the > checkout, there is a file foo/baz. Does the untracked cache need > to get updated? > > Unfortunately, the untracked cache is very unlikely to make it to > the top of my priority list any time soon, so I won't be able to > correct this test (and, if necessary, correct the code). But I > would strongly suggest that the test be improved before this code > is merged. > > Thanks for CCing me. I will try to find time to tweak what was sent to the list here to reflect your affiliations better, but marked with DONTMERGE waiting for the necessary updates you mentioned above, so that this change is not forgotten. It may turn out to be that copying from src to dst like the patch does is all that is needed, or the cache may need further invalidation when the copying happens, and I haven't got a good feeling that anybody who are familiar with the codepath vetted the correctness from seeing the discussion from sidelines (yet). Thanks.
RE: [PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
Can you actually keep the email address as my Twopensource one? I want to make sure that Twitter, my employer at the time, gets credit for this work (just as I want to make sure that my current employer, Two Sigma, gets credit for my current work). Please feel free to add Signed-off-by: David Turner <dtur...@twosigma.com> in case that makes tracking easier. Thanks. WRT the actual patch, I want to note that past me did not do a great job here. The tests do not correctly check that the post-checkout untracked cache is still valid after a checkout. For example, let's say that previously, the directory foo was entirely untracked (but it contained a file bar), but after the checkout, there is a file foo/baz. Does the untracked cache need to get updated? Unfortunately, the untracked cache is very unlikely to make it to the top of my priority list any time soon, so I won't be able to correct this test (and, if necessary, correct the code).But I would strongly suggest that the test be improved before this code is merged. Thanks for CCing me. > -Original Message- > From: Christian Couder [mailto:christian.cou...@gmail.com] > Sent: Monday, May 8, 2017 6:12 AM > To: Johannes Schindelin <johannes.schinde...@gmx.de> > Cc: git <git@vger.kernel.org>; Junio C Hamano <gits...@pobox.com>; Nguyễn > Thái Ngọc Duy <pclo...@gmail.com>; Ben Peart <benpe...@microsoft.com>; > David Turner <david.tur...@twosigma.com> > Subject: Re: [PATCH 0/1] Preserve the untracked cache across checkout, reset > -- > hard, etc > > (Adding Dave in Cc as it looks like he is involved.) > > On Mon, May 8, 2017 at 11:41 AM, Johannes Schindelin > <johannes.schinde...@gmx.de> wrote: > > I recently sent out a request for assistance, after noticing that the > > untracked cache is simply thrown away after operations such as `git > > checkout` or `git reset --hard`: > > > > http://public-inbox.org/git/alpine.DEB.2.20.1705031202470.3480@virtual > > box/ > > > > Duy responded with some high-level reasoning that it should be > > possible to simply reuse the untracked cache data structure in the new > > index, as he had a gut feeling that "we do invalidation right". > > > > I did not have time to back that up by a thorough analysis of the > > code, but it turns out that it is unnecessary: Ben Peart pointed me to > > a patch of Dave Turner's that was submitted as part of the watchman > > series, addressing the very issue about which I was concerned. > > > > And I trust Dave to have validated the idea that the untracked cache > > invalidation "is done right" even when we simply move the pointer to a > > different index_state struct than originally. > > > > Seeing as the untracked cache being dropped unceremoniously when it > > should not be dropped, in a surprising number of operations, I think > > it is a sensible change, and important, too, and independent enough > > from the watchman patches to merit being separated out and applied > > pretty soon. > > > > So what I did was simply to drop the two lines from this patch that > > referred to index_state fields added by Dave's watchman patch series. > > > > Please do not mistake this for a sign that I am disinterested in > > watchman support, far from it... stay tuned ;-) > > > > Oh, and I adjusted Dave's email address. Dave, is that okay? > > > > As we are in a feature freeze phase, I was debating whether to send > > out this patch now or later. > > > > Having thought about it for quite a bit, I am now convinced that this > > patch fixes a bug in the untracked cache feature that is so critical > > as to render it useless: if you > > > > - have to switch between branches frequently, or > > - rebase frequently (which calls `git reset --hard`), or > > - stash frequently (which calls `git reset --hard`), > > > > it is as if you had not enabled the untracked cache at all. Even > > worse, Git will do a ton of work to recreate the untracked cache and > > to store it as an index extension, *just* to throw the untracked away in the > end. > > > > > > David Turner (1): > > unpack-trees: preserve index extensions > > > > cache.h | 1 + > > read-cache.c | 6 ++ > > t/t7063-status-untracked-cache.sh | 22 ++ > > unpack-trees.c| 1 + > > 4 files changed, 30 insertions(+) > > > > > > base-commit: 4fa66c85f11bc5a541462ca5ae3246aa0ce02e74 > > Published-As: > > https://github.com/dscho/git/releases/tag/preserve-untracked-cache-v1 > > Fetch-It-Via: git fetch https://github.com/dscho/git > > preserve-untracked-cache-v1 > > > > -- > > 2.12.2.windows.2.800.gede8f145e06 > >
Re: [PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
(Adding Dave in Cc as it looks like he is involved.) On Mon, May 8, 2017 at 11:41 AM, Johannes Schindelinwrote: > I recently sent out a request for assistance, after noticing that the > untracked cache is simply thrown away after operations such as > `git checkout` or `git reset --hard`: > > http://public-inbox.org/git/alpine.DEB.2.20.1705031202470.3480@virtualbox/ > > Duy responded with some high-level reasoning that it should be possible > to simply reuse the untracked cache data structure in the new index, as > he had a gut feeling that "we do invalidation right". > > I did not have time to back that up by a thorough analysis of the code, > but it turns out that it is unnecessary: Ben Peart pointed me to a patch > of Dave Turner's that was submitted as part of the watchman series, > addressing the very issue about which I was concerned. > > And I trust Dave to have validated the idea that the untracked cache > invalidation "is done right" even when we simply move the pointer to a > different index_state struct than originally. > > Seeing as the untracked cache being dropped unceremoniously when it > should not be dropped, in a surprising number of operations, I think it > is a sensible change, and important, too, and independent enough from > the watchman patches to merit being separated out and applied pretty > soon. > > So what I did was simply to drop the two lines from this patch that > referred to index_state fields added by Dave's watchman patch series. > > Please do not mistake this for a sign that I am disinterested in > watchman support, far from it... stay tuned ;-) > > Oh, and I adjusted Dave's email address. Dave, is that okay? > > As we are in a feature freeze phase, I was debating whether to send out > this patch now or later. > > Having thought about it for quite a bit, I am now convinced that this > patch fixes a bug in the untracked cache feature that is so critical as > to render it useless: if you > > - have to switch between branches frequently, or > - rebase frequently (which calls `git reset --hard`), or > - stash frequently (which calls `git reset --hard`), > > it is as if you had not enabled the untracked cache at all. Even worse, > Git will do a ton of work to recreate the untracked cache and to store > it as an index extension, *just* to throw the untracked away in the end. > > > David Turner (1): > unpack-trees: preserve index extensions > > cache.h | 1 + > read-cache.c | 6 ++ > t/t7063-status-untracked-cache.sh | 22 ++ > unpack-trees.c| 1 + > 4 files changed, 30 insertions(+) > > > base-commit: 4fa66c85f11bc5a541462ca5ae3246aa0ce02e74 > Published-As: > https://github.com/dscho/git/releases/tag/preserve-untracked-cache-v1 > Fetch-It-Via: git fetch https://github.com/dscho/git > preserve-untracked-cache-v1 > > -- > 2.12.2.windows.2.800.gede8f145e06 >
[PATCH 0/1] Preserve the untracked cache across checkout, reset --hard, etc
I recently sent out a request for assistance, after noticing that the untracked cache is simply thrown away after operations such as `git checkout` or `git reset --hard`: http://public-inbox.org/git/alpine.DEB.2.20.1705031202470.3480@virtualbox/ Duy responded with some high-level reasoning that it should be possible to simply reuse the untracked cache data structure in the new index, as he had a gut feeling that "we do invalidation right". I did not have time to back that up by a thorough analysis of the code, but it turns out that it is unnecessary: Ben Peart pointed me to a patch of Dave Turner's that was submitted as part of the watchman series, addressing the very issue about which I was concerned. And I trust Dave to have validated the idea that the untracked cache invalidation "is done right" even when we simply move the pointer to a different index_state struct than originally. Seeing as the untracked cache being dropped unceremoniously when it should not be dropped, in a surprising number of operations, I think it is a sensible change, and important, too, and independent enough from the watchman patches to merit being separated out and applied pretty soon. So what I did was simply to drop the two lines from this patch that referred to index_state fields added by Dave's watchman patch series. Please do not mistake this for a sign that I am disinterested in watchman support, far from it... stay tuned ;-) Oh, and I adjusted Dave's email address. Dave, is that okay? As we are in a feature freeze phase, I was debating whether to send out this patch now or later. Having thought about it for quite a bit, I am now convinced that this patch fixes a bug in the untracked cache feature that is so critical as to render it useless: if you - have to switch between branches frequently, or - rebase frequently (which calls `git reset --hard`), or - stash frequently (which calls `git reset --hard`), it is as if you had not enabled the untracked cache at all. Even worse, Git will do a ton of work to recreate the untracked cache and to store it as an index extension, *just* to throw the untracked away in the end. David Turner (1): unpack-trees: preserve index extensions cache.h | 1 + read-cache.c | 6 ++ t/t7063-status-untracked-cache.sh | 22 ++ unpack-trees.c| 1 + 4 files changed, 30 insertions(+) base-commit: 4fa66c85f11bc5a541462ca5ae3246aa0ce02e74 Published-As: https://github.com/dscho/git/releases/tag/preserve-untracked-cache-v1 Fetch-It-Via: git fetch https://github.com/dscho/git preserve-untracked-cache-v1 -- 2.12.2.windows.2.800.gede8f145e06