Web lists-archives.com

Bloom Filters (was Re: We should add a "git gc --auto" after "git clone" due to commit graph)

(Changing title to reflect the new topic.)

On 10/8/2018 11:08 PM, Jeff King wrote:
On Mon, Oct 08, 2018 at 02:29:47PM -0400, Derrick Stolee wrote:

There are two questions that I was hoping to answer by looking at your code:
1. How do you store your Bloom filter? Is it connected to the commit-graph
and split on a commit-by-commit basis (storing "$path" as a key), or is it
one huge Bloom filter (storing "$commitid:$path" as key)?
I guess you've probably thought all of this through for your
implementation, but let me pontificate.

I'd have done it as one fixed-size filter per commit. Then you should be
able to hash the path keys once, and apply the result as a bitwise query
to each individual commit (I'm assuming that it's constant-time to
access the filter for each, as an index into an mmap'd array, with the
offset coming from a commit-graph entry we'd be able to look up anyway).

You're right that we want to hash the path a constant number of times. Add in that we want to re-use information already serialized when updating the file, and we see that having a commit-by-commit Bloom filter is a good idea. Using (commit, path) pairs requires lots of re-hashing, repeated work when extending the filter, and poor locality when evaluating membership of a single key.

The nice thing is that you can generate k 32-bit hash values based on two 32-bit hash values that are "independent enough" (see [1]). We used Murmur3 with two different seed values to generate hashes a & b, then used the arithmetic progression a, a + b, a + 2b, ..., a + (k-1)b as our k hash values. These can be computed up front and then dropped into any size filter using a simple modulo operation. This allows flexible sizes in our filters.

I don't think fixed size filters are a good idea, and instead would opt for flex-sized filters with a maximum size. The typical parameters use 7 hash functions and a filter size of (at least) 10 bits per entry. For most commits (say 60-70%), 256 bits (32 bytes) is enough. Using a maximum of 512 bytes covers 99% of commits. We will want these bounds to be configurable via config. If we had a fixed size, then we either make it too small (and don't have sufficient coverage of commits) or too large (and waste a lot of space on the commits that change very little).

We can store these flex-sized filters in the commit-graph using two columns of data (two new optional chunks):

* Bloom filter data: stores the binary data of each commit's Bloom filter, concatenated together in the same order as the commits appear in the commit-graph.

* Bloom filter positions: The ith position of this column stores the start of the (i+1)th Bloom filter (The 0th filter starts at byte 0). A Bloom filter of size 0 is intended to mean "we didn't store this filter because it would be too large". We can compute the lengths of the filter by inspecting adjacent values.

In order to be flexible, we will want to encode some basic information into the Bloom filter data chunk, such as a tuple of (hash version, num hash bits, num bits per entry). This allows us to change the parameters in config but still be able to read a serialized filter. Here I assume that all filters share the same parameters. The "hash version" here is different than the_hash_algo, because we don't care about cryptographic security, only a uniform distrubution (hence, Murmur3 is a good, fast option).

[1] https://web.archive.org/web/20090131053735/http://www.eecs.harvard.edu/~kirsch/pubs/bbbf/esa06.pdf

I think it would also be easier to deal with maintenance, since each
filter is independent (IIRC, you cannot delete from a bloom filter
without re-adding all of the other keys).

The obvious downside is that it's O(commits) storage instead of O(1).
It would always be O(changes), as the Bloom filter needs to grow in size as the number of entries grows.
Now let me ponder a bit further afield. A bloom filter lets us answer
the question "did this commit (probably) touch these paths?". But it
does not let us answer "which paths did this commit touch?".

That second one is less useful than you might think, because we almost
always care about not just the names of the paths, but their actual
object ids. Think about a --raw diff, or even traversing for
reachability (where if we knew the tree-diff cheaply, we could avoid
asking "have we seen this yet?" on most of the tree entries). The names
alone can make that a bit faster, but in the worst case you still have
to walk the whole tree to find their entries.

But there's also a related question: how do we match pathspec patterns?
For a changed path like "foo/bar/baz", I imagine a bloom filter would
mark all of "foo", "foo/bar", and "foo/bar/baz". But what about "*.c"? I
don't think a bloom filter can answer that.

The filter needs to store every path that would be considered "not TREESAME". It can't store wildcards, so you would need to evaluate the wildcard and test all of those paths individually (not a good idea).

At least not by itself. If we imagine that the commit-graph also had an
alphabetized list of every path in every tree, then it's easy: apply the
glob to that list once to get a set of concrete paths, and then query
the bloom filters for those. And that list actually isn't too big. The
complete set of paths in linux.git is only about 300k gzipped (I think
that's the most relevant measure, since it's an obvious win to avoid
repeating shared prefixes of long paths).
As you mention below, we would actually want a list of "every path that has ever appeared in the repo".
Imagine we have that list. Is a bloom filter still the best data
structure for each commit? At the point that we have the complete
universe of paths, we could give each commit a bitmap of changed paths.
That lets us ask "did this commit touch these paths" (collect the bits
from the list of paths, then check for 1's), as well as "what paths did
we touch" (collect the 1 bits, and then index the path list).  Those
bitmaps should compress very well via EWAH or similar (most of them
would be huge stretches of 0's punctuated by short runs of 1's).

I'm not convinced we would frequently have runs of 1's, and the bitmap would not compress much better than simply listing the positions. For example, a path "foo/bar" that resolves to a tree would only start a run if the next changes are the initial section of entries in that tree (sorted lexicographically) such as "foo/bar/a, foo/bar/b". If we deepen into a tree, then we will break the run of 1's unless we changed every path deeper than that tree.

So that seems promising to me (or at least not an obvious dead-end). I
do think maintenance gets to be a headache, though. Adding new paths
potentially means reordering the bitmaps, which means O(commits) work to
"incrementally" update the structure. (Unless you always add the new
paths at the end, but then you lose fast lookups in the list; that might
be an acceptable tradeoff).

And finally, there's one more radical option: could we actually store a
real per-commit tree-diff cache? I.e., imagine that each commit had the
equivalent of a --raw diff easily accessible, including object ids. That
would allow:

   - fast pathspec matches, including globs

   - fast --raw output (and faster -p output, since we can skip the tree

   - fast reachability traversals (we only need to bother to look at the
     objects for changed entries)

where "fast" is basically O(size of commit's changes), rather than
O(size of whole tree). This was one of the big ideas of packv4 that
never materialized. You can _almost_ do it with packv2, since after all,
we end up storing many trees as deltas. But those deltas are byte-wise
so it's hard for a reader to convert them directly into a pure-tree
diff (they also don't mention the "deleted" data, so it's really only
half a diff).

So let's imagine we'd store such a cache external to the regular object
data (i.e., as a commit-graph entry). The "log --raw" diff of linux.git
has 1.7M entries. The paths should easily compress to a single 32-bit
integer (e.g., as an index into a big path list). The oids are 20 bytes.
Add a few bytes for modes. That's about 80MB. Big, but not impossibly
so. Maybe pushing it for true gigantic repos, though.

Above, I mentioned my gut reaction that storing a "changed path bitmap" per commit would not compress well. That puts that implementation very close to the one you suggest here (except we also store the OID changes).

I just want to compare your 80MB here to ~4MB it would take to store those changed paths in Bloom filters (10 bits per entry -> ~2MB, but adding some slop for the commit-by-commit storage).

Those numbers are ignoring merges, too. The meaning of "did this commit
touch that path" is a lot trickier for a merge commit, and I think may
depend on context. I'm not sure how even a bloom filter solution would
handle that (I was assuming we'd mostly punt and let merges fall back to
opening up the trees).

My solution here is to always store the list of paths changed against the first parent. If we evaluate TREESAME against our first parent while computing simplified file history, then we continue along first-parent history. It is possible to store filters for every parent, but I don't recommend it. The merge commit will typically have many more change paths against the second parent, since the second parent is usually bringing in a small change done by few to catch up to the work done in parallel by many. Those diffs will frequently run over our limit.

Phew. That was a lot. I don't want to derail any useful work either of
you is doing. These are just things I've been thinking over (or even in
some cases experimenting with), and I think it's worth laying all the
options on the table. I won't be surprised if you'd considered and
rejected any of these alternate approaches, but I'd be curious to hear
the counter-arguments. :)

This is a good discussion to have, since the commit-graph feature is getting to a stable place. We still have ongoing algorithm work with generation numbers, but this Bloom filter discussion (and implementation) can happen in parallel.