1
This document contains notes about the design for groupcompress, replacement
2
VersionedFiles store for use in pack based repositories. The goal is to provide
3
fast, history bounded text extraction.
8
The goal: Much tighter compression, maintained automatically. Considerations
9
to weigh: The minimum IO to reconstruct a text with no other repository
10
involved; The number of index lookups to plan a reconstruction. The minimum
11
IO to reconstruct a text with another repositories assistance (affects
12
network IO for fetch, which impacts incremental pulls and shallow branch
18
Each delta is individually compressed against another text, and then entropy
19
compressed. We index the pointers between these deltas.
21
Solo reconstruction: Plan a readv via the index, read the deltas in forward
22
IO, apply each delta. Total IO: sum(deltas) + deltacount*index overhead.
23
Fetch/stacked reconstruction: Plan a readv via the index, using local basis
24
texts where possible. Then readv locally and remote and apply deltas. Total IO
25
as for solo reconstruction.
30
Reasonable sizes 'amount read' from remote machines to reconstruct an arbitrary
31
text: Reading 5MB for a 100K plain text is not a good trade off. Reading (say)
32
500K is probably acceptable. Reading ~100K is ideal. However, it's likely that
33
some texts (e.g NEWS versions) can be stored for nearly-no space at all if we
34
are willing to have unbounded IO. Profiling to set a good heuristic will be
35
important. Also allowing users to choose to optimise for a server environment
36
may make sense: paying more local IO for less compact storage may be useful.
41
Index scatter gather IO. Doing hundreds or thousands of index lookups is very
42
expensive, and doing that per file just adds insult to injury.
44
Partioned compression amongst files.
46
Scatter gather IO when reconstructing texts: linear forward IO is better.
51
Merges combine texts from multiple versions to create a new version. Deltas
52
add new text to existing files and remove some text from the same. Getting
53
high compression means reading some base and then a chain of deltas (could
54
be a tree) to gain access to the thing that the final delta was made against,
55
and that delta. Rather than composing all these deltas, we can just just
56
perform the final diff against the base text and the serialised invidual
57
deltas. If the diff algorithm can reuse out of order lines from previous
58
texts (e.g. storing AB -> BA as pointers rather than delete and add, then
59
the presence of any previously stored line in a single chain can be reused.
60
One such diff algorithm is xdelta, another reasonable one to consider is
61
plain old zlib or lzma. We could also use bzip2. One advantage of using
62
a generic compression engine is less python code. One advantage of
63
preprocessing line based deltas is that we reduce the window size for the
64
text repeated within lines, and that will help compression by a simple
65
entropy compressor as a post processor.
66
lzma appears fantastic at compression - 420MB of NEWS files down to 200KB.
67
so window size appears to be a key determiner for efficiency.
72
Very big objects - no delta. I plan to kick this in at 5MB initially, but
73
once the codebase is up and running, we can tweak this to
75
Very small objects - no delta? If they are combined with a larger zlib object
76
why not? (Answer: because zlib's window is really small)
78
Other objects - group by fileid (gives related texts a chance, though using a
79
file name would be better long term as e.g. COPYING and COPYING from different
80
projects could combine). Then by reverse topological graph(as this places more
81
recent texts at the front of a chain). Alternatively, group by size, though
82
that should not matter with a large enough window.
83
Finally, delta the texts against the current output of the compressor. This is
84
essentially a somewhat typed form of sliding window dictionary compression. An
85
alternative implementation would be to just use zlib, or lzma, or bzip2
88
Unfortunately, just using entropy compression forces a lot of data to be output
89
by the decompressor - e.g. 420MB in the NEWS sample corpus. When we only want
90
a single 55K text thats inefficient. (An initial test took several seconds with
93
The fastest to implement approach is probably just 'diff output to date and add
94
to entropy compressor'. This should produce reasonable results. As delta
95
chain length is not a concern (only one delta to apply ever), we can simply
96
cap the chain when the total read size becomes unreasonable. Given older texts
97
are smaller we probably want some weighted factor of plaintext size.
99
In this approach, a single entropy compressed region is read as a unit, giving
100
the lower bound for IO (and how much to read is an open question - what byte
101
offset of compressed data is sufficient to ensue that the delta-stream contents
102
we need are reconstructable. Flushing, while possible, degrades compression(and
103
adds overhead - we'd be paying 4 bytes per record guaranteed). Again - tests
106
A nice possibility is to output mpdiff compatible records, which might enable
107
some code reuse. This is more work than just diff (current_out, new_text), so
108
can wait for the concept to be proven.
110
Implementation Strategy
111
+++++++++++++++++++++++
113
Bring up a VersionedFiles object that implements this, then stuff it into a
114
repository format. zlib as a starting compressor, though bzip2 will probably