On Mon, May 4, 2020 at 3:31 PM Johannes Schindelin via GitGitGadget
[off-list ref] 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

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.

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.

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.

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. 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.

Here is my counterpoint to your proposal:

Reftable was developed outside of git-core. 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.

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.

Could you explain to me how you would qualify a reftable-in-git
library against the JGit implementation?

-- 
Han-Wen Nienhuys - Google Munich
I work 80%. Don't expect answers from me on Fridays.
--
Google Germany GmbH, Erika-Mann-Strasse 33, 80636 Munich
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg
Geschäftsführer: Paul Manicle, Halimah DeLaine Prado