Hi Han-Wen, On Mon, 4 May 2020, Han-Wen Nienhuys wrote: > On Mon, May 4, 2020 at 3:31 PM Johannes Schindelin via GitGitGadget > <gitgitgadget@xxxxxxxxx> wrote: > > > > These patches are not intended to be complete, not by any stretch of > > imagination. They are just enough to get the CI run to pass, even in the > > Windows-specific parts. > > Thanks for the investigation. I had begun to setup a Windows > installation, but you beat me to it. > > Here are the two fixes as I applied them. > > https://github.com/google/reftable/commit/201f21aa39f993c970ef4d19122fedd5b91e716d > https://github.com/google/reftable/commit/18b041472306b31813dbd6408f6a61daaba60013 If you ever needed _any_ kind of proof that the current setup with that monster patch and a totally separate "upstream" does simply _not_ work, it is these two commits. _Every_ bug fix originating from git.git would have to manually port the fixes because of the vendor copy, just like you did, losing all provenance apart from a meager "Thanks to" (which mentions just the bug report forgetting about the actually-tested patch). But I don't get the impression that my objections are heard, so I might just as well stop talking about it. Please do not misinterpret my silence about your patches as any indication that I am happy with `hn/reftable`: I am not. Should Junio accept them without drastic changes (which I hope he will not), I expect this to turn into a complete maintenance nightmare. Besides, it is a bit disappointing to see only those two patches. I made very clear in my patch series that I got the _distinct_ impression that those two bug fixes were just the tip of the iceberg: there are most likely a lot more bugs waiting to be discovered where a file is not closed when it is no longer needed (or related: memory not being released as soon as it can be released). I expect _many_ more similar issues to lurk in the code base. > > As I mentioned elsewhere, I would much prefer for hn/reftable to not > > re-invent get_be*(), struct strbuf, struct string_list, struct lock_file > > etc. > > > > However, in the context of the test failures, I do not think that this would > > have made a big difference. Apart from the unportable constructs, and from > > the "delete/rename while there is still a handle on the file" issues, it > > would appear that one big reason why it was so hard to debug and fix the > > test is the recursive complexity of the data structures. > > > > To elaborate on that: struct reftable_stack has an attribute called merged > > that has an array of struct reftable_reader * (confusingly called "stack"). > > Each of these readers has an attribute of type struct reftable_block_source > > that uses a vtable and an opaque pointer to abstract three types of block > > sources: file, slice (which is apparently unused, i.e. it is apparently just > > dead weight for now) and malloc. > > All of the unittests use slices to read and write single reftables > in-memory. > > > The malloc block source is used to swap out a block coming from the > file directly, with zlib uncompressed data (log blocks are zlib > compressed). I think there should be a mmap block source too, at some > point. If that was intended to convince me that I am wrong in my assessment of the unnecessary complexity of libreftable, I am afraid it did not. Only code should be included in the patch adding libreftable to git.git which is actually needed in git.git. Everything above that is just unnecessary complexity. > > I am not sure that this abstraction fest serves us well here. > > > > Quite honestly, I would have expected the packed_git data structure to be > > imitated more closely, as it strikes me as similar in purpose, and it has > > seen a ton of testing over the past decade and a half. I could not recognize > > that design in the reftable, though. > > > > It is quite obvious, of course, that the code tries to imitate the > > object-oriented nature of the Go code, but it seems quite obvious from my > > difficulties to address the problem where stack_compact_range() would try to > > delete stale reftable files (without even so much as a warning when the > > files could not be deleted!) without releasing all file handles to all > > reftable files, even the ones that do not need to be deleted. To be smarter > > about this, the code has to know more about the nature of the block source > > than the abstraction layer suggests. It has to know that a block source > > refers to a file, and that that file is marked for deletion. My heavy-handed > > work-around, even if it works, is not really a good solution, but it is a > > testament that there is a lot of opportunity to simplify the code > > drastically while at the same time simplifying the design, too. > > If the code tries to delete a file while it is open, that is a > testament to the fact that I haven't written code for Windows for many > years now. It is not a testament to any fundamental problems with the > library design. To the contrary. The library design neglects taking good care of resources (such as closing file descriptors when they are clearly no longer needed). C code _has_ to be careful with resources. By the way, I am not really trying to convince you of this. That would be the wrong direction: you have to convince _me_, the reviewer, that your patches do the right thing, in the right way. So far, I am unconvinced. Even worse: my debugging session two days ago convinced me that there _is_ unnecessary complexity, and unclear resource management. Based on how much experience I have with debugging and developing in unfamiliar code bases, this means quite a bit. I am not a junior intern who gets easily overwhelmed by code they haven't seen before. When I point out the unnecessary complexity of having a vtable for a mere two "subclasses", of course, as an open source contributor you are free to dismiss this without even so much as trying to refute it. Just like the Git project is free to reject contributions when the reviewer comments are not taken seriously. > The library which you are complaining of weighs in at about 7500 lines > of code (excluding tests), which compares pretty well to the original > JGit code which is now ~6500 lines of code. I don't think there is > room to simplify it further, and I say this as a person who has > significant experience with this format. > > If you wish to prove me wrong, you can send me patch. Until then, I > don't buy your arguments. Again, it is not so much a question whether _you_ buy _my_ arguments (and at the same time, it is somewhat shocking that you dismiss that it took me two hours to figure out a single bug, just because the code and the data structures are unnecessarily complex/abstract). It is more a question whether you can convince the reviewers that your code is correct. In short: I am not selling anything here that you need to buy. You are trying to sell me something, and I do not like the shape, and I like even less how easily you seem to brush aside my explanations how this would need to be improved to make it viable in Git's source code. > > I know you have been putting a lot of effort into this library, so I feel a > > bit bad about saying the following: The hn/reftable patches will need > > substantial work before we can merge it with confidence. Part of the reason > > why it is so hard to review the reftable patches is that they intentionally > > refuse to integrate well within Git's source code, such as (re-)implementing > > its own fundamental data structures, intentionally using a totally different > > coding style, and it concerns me that the stated requirement for bug fixes > > is to treat Git's source code as a downstream of the actual project. I am > > not too bad a C developer and would consider myself competent in debugging > > issues, even hard ones, in Git, and yet... it was really hard to wade > > through the layers of abstraction to figure out where the file handles > > should be closed that were opened and prevented deleting/renaming files. > > > > At this point, I don't feel that it makes sense to keep insisting on having > > this in a separate library. The only other user of that library will be > > libgit2, and I really think that given libgit2's heritage, it won't be a > > problem to adapt the code after it stabilized in git.git (and since libgit2 > > treats git.git as upstream for the libxdiff code, it won't be a problem to > > do the same for the reftable code, too). I believe that the best course of > > action is to reuse the data structures libgit.a provides, and to delete the > > re-implementations in reftable/. Only then can we start working effectively > > on refactoring the code to simplify the data structures in order to clarify > > resource usage (which was the root cause for the bugs I fixed, although I am > > sure that there are way more lurking in there, hidden by the fact that the > > code is not covered thoroughly by our tests). > > I'm happy to change the library to use more common primitives that are > shared between libgit2 and git. I never talked about primitives that are shared between libgit2. I talked about data structures and helpers that already exist in Git's source code. And I said that I am fairly convinced that it won't be hard to adapt the code to libgit2 after it enters git.git. I am sorry if I did not make myself clear. It is a bit unreasonable to reject a review on the basis that git.git should accept cruft that _might_ be required to support libgit2 better. It would be very unreasonable to accept a patch series that demands to include code in git.git that it does not need just because a separate project might need that code. And if you truly care about libgit2 support, you will include the libgit2 developers in the development of the reftable support, not throw a nominally finalized library at them that might, or might not integrate well into libgit2. > Could you point them out to me? Note that the basics that you are > complaining of (put_be*, strbuf vs slice routines, etc.) constitute > around 700 lines of code. It's not going to make an appreciable > difference in complexity. Was I unclear? This additional code _makes it harder to review the monster patch_. The complexity _will_ need to be reduced. In addition. But that cannot be the first step: in order to integrate reasonably well with Git's source code, it will _have_ to stop duplicating code. You cannot possibly deny that you reimplemented your own string manipulation library in libreftable. Such code adds to the maintenance burden, and in this instance (and in other, similar instances) for no good reason at all because libgit.a already has code for that. Without integrating the code better in git.git, there is no sense to even start reducing the complexity. Both will be necessary. > Here is my counterpoint to your proposal: > > Reftable was developed outside of git-core. That was your decision, not mine. I merely pointed out that it would have been wiser to develop it within. And I offered that it is undesirable to take `hn/reftable` as long as it looks very much like it _wants_ to stay out of core Git. I mean, what's the point of accepting code into our code base that does not really want to live there? > If you feel this series with its giant patch is too much to review, you > can have a look at the incremental story of how it came to be, both in > the history and code reviews of the JGit project, and in the commit > history of > https://github.com/google/reftable. So now you want to place _even more_ of a burden on reviewers? Seriously? > The idea to put the code into git-core, and especially your proposed > (but unsubstantiated) plans to "simplify" its design, make me very > worried about interoperability with the JGit reference implementation. The JGit reference implementation is written in Java. In an object-oriented language with its own idiosyncracies demanded by the design of said language (the lengths to which Shawn had to go to make it performant are recorded in the commit history). It will be better to have a clean-room implementation in C that does not suffer from transpiling artifacts. About ways to simplify its design: I mentioned the layers of abstraction. I mentioned the design of `struct packed_git`, which could (and should) serve as an example to follow. If those points weren't ignored, I would mention other suggestions, such as how the code builds file contents via `slice_append_string()` only to write them out in one go, when the contents could have been written much more simply via `fprintf()` instead. And that is just _one_ more example how complexity could be reduced, very, very easily. There's much more. Would there be a point of me writing these up? So far, I don't get that impression. All I get is "I don't buy this". That is unnecessarily frustrating: code review is already thankless (as much as it is necessary!) as it is. > Could you explain to me how you would qualify a reftable-in-git > library against the JGit implementation? Of course: by a comprehensive set of unit/regression tests, including pre-generated data that the implementation must read and parse correctly. The fact that keeping the code close to its non-C origins resulted in long-undetected segmentation faults makes me rather certain the idea of keeping the code close to said reference implementation is overrated: it does not build the confidence in the code that you think it does. Not after discovering those segmentation faults, it doesn't. At the end of the day, a comprehensive set of unit/regression tests will be required for implementing reftable support in other projects, anyway, in particular for projects that cannot, or do not want to, rely on your libreftable implementation, such as isomorphic-git or Dulwich. Ciao, Johannes
