Merge lp:~ian-clatworthy/bzr/faster-dirstate-saving into lp:~bzr/bzr/trunk-old

Many commonly used commands like status, add and commit update the dirstate, triggering a dirstate serialisation & save. On huge trees like OpenOffice, this is slower than it needs to be. In particular, 'xxx status file' takes 0.4 seconds in hg and 1.0 seconds in bzr and a good percentage of the difference is due to the time we take to serialise the new dirstate.

This branch is an experiment/RFC in fixing that. It drops the time for 'bzr status file' by 30-35% down to 0.65-0.70 seconds. It does that by remembering the serialised form of entries and only re-serialising entries that are known to be changed. Right now, this smart remembering of what's changed is only effectively implemented for status, though the internal API is in place for extending that to other use cases.

Of course, there are other ways of skinning this cat. One option is to write a pyrex serialiser. That ought to be fast but it still doesn't solve the root problem: serialisation time is O(size-of-tree) currently because we only keep a modified vs unmodified flag at the whole-of-dirstate level. Another option is to append 'overlays' to the dirstate file, i.e. entries which have been added or changed vs the base entries. Deletes or renames would trigger a full clean write but the common cases of add and/or change would just append entries. That's non-trivial but potentially very fast.

More broadly, I think the important thing to be begin recording the changes as this patch allows. So my current thoughts are that we ought to start with this patch, make the changes to enable smarter recording for add and commit, and built from there. At any point, we can separately do a pyrex serialiser and it will complement this work.

Having said all that, dirstate is my least favourite part of the Bazaar code base: indexing into tuples using magic integers may be fast but it sucks from an understandability perspective (vs objects + attributes). There are people far more qualified than I to say how this ought to proceed and to write the code, but they're rather busy tackling other things. Regardless, it's been a good exercise for me in getting dirstate paged into my head for other work I'm doing. It's a step forward but I can't definitively say it's in the right direction.

Thoughts?

Brief thought: perhaps simply don't bother updating the dirstate file if the number of changes is very small? (Or do we already do this?)

> Brief thought: perhaps simply don't bother updating the dirstate file if the
> number of changes is very small? (Or do we already do this?)

Poolie has agreed with this so I'll go ahead and resubmit accordingly. The updated benchmark time for 'bzr status file' on OOo is 0.5 seconds, down from 1.0 second.

We just need to agree on a # of changes below which it's not worth saving. I suggest we should save for 3 or more changes - so we skip iff 2 or less files are changed. If someone has a better number, let me know.

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Ian Clatworthy wrote:
>> Brief thought: perhaps simply don't bother updating the dirstate file if the
>> number of changes is very small? (Or do we already do this?)
>
> Poolie has agreed with this so I'll go ahead and resubmit accordingly. The updated benchmark time for 'bzr status file' on OOo is 0.5 seconds, down from 1.0 second.
>
> We just need to agree on a # of changes below which it's not worth saving. I suggest we should save for 3 or more changes - so we skip iff 2 or less files are changed. If someone has a better number, let me know.

A guideline is: "Don't save if it would cost as much data as you avoid
reading".

The cost of not updating the dirstate is having to re-read the file
who's sha is now different. So if the dirstate is 1MB, and the file is
10k, then you should not update if there are 10 files that are changed.

If you *want* to look at text_size and compare it to len(lines) we could
do that. Otherwise I'm fine with a heuristic of < 1% of the tree, or
something like that.

John
=:->
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (Cygwin)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iEYEARECAAYFAkoepA4ACgkQJdeBCYSNAANrVwCgwjbMhA60vNrtlpbciGdFw9pp
Y6MAn1FfgUIcKOSm5BjgXP5EjKJtbaVS
=i1ML
-----END PGP SIGNATURE-----

2009/5/28 John A Meinel <email address hidden>:
> A guideline is: "Don't save if it would cost as much data as you avoid
> reading".
>
> The cost of not updating the dirstate is having to re-read the file
> who's sha is now different. So if the dirstate is 1MB, and the file is
> 10k, then you should not update if there are 10 files that are changed.
>
> If you *want* to look at text_size and compare it to len(lines) we could
> do that. Otherwise I'm fine with a heuristic of < 1% of the tree, or
> something like that.

It's possibly a bit more work, but as I commented on the bug, I'd
actually like to try turning off writing it altogether for logical
readonly operations.

--
Martin <http://launchpad.net/~mbp/>

This updated merge proposal makes it configurable as to how many changes are ok before the dirstate should be saved. As Martin requested, there's the ability to turn off saving altogether. After this lands, we can use that to implement various policies (e.g. don't save on a logical read-only operation) as we see fit.

This drops 'bzr st [file]' down by 0.5+ seconds on OOo. Nice. It doesn't yet make a difference for add or commit. Those are small follow-on patches after we're happy with this one.

I'm no dirstate expert, so depending on how confident you are with this patch you might like to wait for another review before merging.

Overall, I this change seems ok to me. It deserves careful benchmarking of course, but I know you've been doing that :)

It's interesting to me that _build_line_cache actually re-reads the lines from the file. I had expected that maybe we'd keep the lines around from the original read, but perhaps this is just as good. If we mmap'ed the dirstate then certainly re-reading from the mmap rather than keeping a copy would clearly make sense.

+ :param worth_saving_limit: when the exact number of changed entries
+ is known, only bother saving the dirstate if more than this
+ count of entries have changed. -1 means never save.

I'd personally prefer "None means never save", but that's not a big deal.

I believe the dirstate can record changes to tree state, not just cache stat info... so is there a risk with this patch that e.g. "bzr mv foo bar" will go unrecorded because it's under the _worth_saving_limit? I understand the value in not updating the cache for only a small number of changes, but AFAIK we always want to update the dirstate file when actual state is changed. I think the header_too=True flag is used to make sure this works, but that seems like a pretty indirect way of expressing the concept that this is a non-optional update.

I wonder: if you unconditionally set _worth_saving_limit to something very high (e.g. 1000000) does the test suite still pass? My guess is that except perhaps for one or two hyper-sensitive tests it ought to, and it might make for an interesting sanity check.

+ # We only enable smart saving is it hasn't already been disabled

"is" -> "if"

+ if self._use_smart_saving is not False:
+ self._use_smart_saving = True

So if _use_smart_saving is True, set it to True? Is that right? Perhaps you mean "is not None"?

(If you really are just testing for True/False, it's generally a bad idea to test for True or False with "is". Use "if [not] self._use_smart_saving:" instead, which is also easier to read.)

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Andrew Bennetts wrote:
> Review: Approve
> I'm no dirstate expert, so depending on how confident you are with this patch you might like to wait for another review before merging.
>
> Overall, I this change seems ok to me. It deserves careful benchmarking of course, but I know you've been doing that :)
>
> It's interesting to me that _build_line_cache actually re-reads the lines from the file.

I'm a bit surprised that re-reading the file is significantly faster
than converting a tuple of strings back into a serialized string. If it
is significant, than I'd certainly rather a pyrex extension than reading
from disk...

>
> + :param worth_saving_limit: when the exact number of changed entries
> + is known, only bother saving the dirstate if more than this
> + count of entries have changed. -1 means never save.
>
> I'd personally prefer "None means never save", but that's not a big deal.
>
> I believe the dirstate can record changes to tree state, not just cache stat info...
> so is there a risk with this patch that e.g. "bzr mv foo bar" will go
unrecorded because it's under the _worth_saving_limit? I understand the
value in not updating the cache for only a small number of changes, but
AFAIK we always want to update the dirstate file when actual state is
changed. I think the header_too=True flag is used to make sure this
works, but that seems like a pretty indirect way of expressing the
concept that this is a non-optional update.

The *big* think to watch out for is that there is "IN_MEMORY_MODIFIED"
that happens from a stat cache update, which can be trivially ignored.
And there is "IN_MEMORY_MODIFIED" that happens from "bzr add" which
*must* be written out.

I didn't look at the code to see if Ian clearly distinguished them, if
he has, then we are fine.

John
=:->
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (Cygwin)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iEYEARECAAYFAkoogDcACgkQJdeBCYSNAAPFWQCfeAnQLM2UPq6K+BVsleEy+szw
UIsAoKeT7L7Jybvz8vHUZI32kw9XSmNP
=SWQG
-----END PGP SIGNATURE-----

Andrew Bennetts wrote:
> Review: Approve
>
> It's interesting to me that _build_line_cache actually re-reads the lines from the file. I had expected that maybe we'd keep the lines around from the original read, but perhaps this is just as good. If we mmap'ed the dirstate then certainly re-reading from the mmap rather than keeping a copy would clearly make sense.
>
>
Andrew,

Thanks for the review!

I did carefully benchmark a few approaches here and, as discussed on
IRC, not doing anything unless we absolutely have to gives the best
result for status. For other operations (that don't leverage the ability
to track some changes yet), the best time to read is probably just
before we mutate the dirstate. FWIW, profiling showed file reading time
to be a small factor - it was rebuilding the keys in the right order
that mattered. Before mutation, we simply grab them straight out of the
dirstate. After then, we need to rebuild them from the file content.

> I believe the dirstate can record changes to tree state, not just cache stat info... so is there a risk with this patch that e.g. "bzr mv foo bar" will go unrecorded because it's under the _worth_saving_limit? I understand the value in not updating the cache for only a small number of changes, but AFAIK we always want to update the dirstate file when actual state is changed. I think the header_too=True flag is used to make sure this works, but that seems like a pretty indirect way of expressing the concept that this is a non-optional update.
>
>
With this patch, there's no risk here. When follow-on patches arrive for
operations other than status, I'll need to handle those sorts of cases.

> I wonder: if you unconditionally set _worth_saving_limit to something very high (e.g. 1000000) does the test suite still pass? My guess is that except perhaps for one or two hyper-sensitive tests it ought to, and it might make for an interesting sanity check.
>
>
I tried this during development. Some low level tests fail that "make
change then read disk content". Otherwise, it looked ok to me.

> + # We only enable smart saving is it hasn't already been disabled
>
> "is" -> "if"
>
> + if self._use_smart_saving is not False:
> + self._use_smart_saving = True
>
> So if _use_smart_saving is True, set it to True? Is that right? Perhaps you mean "is not None"?
>
>
As discussed on IRC, I'll make this "if xxx is None".

Once again, thanks for the review.

Ian C.

I'd just like to make it clear that I'm "hesitant" on this patch. I'll admit I haven't looked at it closely, but it seems a bit along the lines of a band-aid for now sort of thing.

Especially given the complexity of keeping multiple inventories in sync...

221 + if self._dirblock_state == DirState.IN_MEMORY_UNMODIFIED:
222 + keys = []
223 + for directory in self._dirblocks:
224 + keys.extend([e[0] for e in directory[1]])
225 + else:

What is the use case for calling _build_line_cache when the current state is IN_MEMORY_UNMODIFIED? Certainly save() is already defined as a no-op in that case. I'd rather avoid having extra code paths that aren't really being used.
I guess maybe it is if the header changes and the blocks don't...

I'll just mention that if you look over the code, the only time _header_state = ...MODIFIED is when _dirblock_state = ... MODIFIED

The only reason it has a separate state flag, is because you can read the header without reading the rest of the dirblocks. (At least to date, certainly the knob exists for that to not be true.)
Though if you consider what is in the header:

1) num parents / parent ids (if this changes, the dirblock state certainly should)
2) ghost parents (see 1)
3) crc (Also indicative of actual content changes)
4) num_entries (if this changes, we better have dirblock changes)

re: safety.

Can I propose that we have an explicit API for just updating the hash
cache (we may already). And that the current state of MODIFIED *always*
refer to having semantic changes rather than hash cache updates.

Then, the hash cache updating code should not alter the dirstate state.
Rather it should just accumulate into a 'hash cache changes counter'.
And save can check that.

This may be very similar to the code. The key thing is that I hope this
will structure things so as to avoid needing to be careful in commands.

-Rob