This document contains notes about the design for groupcompress, replacement VersionedFiles store for use in pack based repositories. The goal is to provide fast, history bounded text extraction.


The goal: Much tighter compression, maintained automatically. Considerations to weigh: The minimum IO to reconstruct a text with no other repository involved; The number of index lookups to plan a reconstruction. The minimum IO to reconstruct a text with another repositories assistance (affects network IO for fetch, which impacts incremental pulls and shallow branch operations).

Current approach

Each delta is individually compressed against another text, and then entropy compressed. We index the pointers between these deltas.

Solo reconstruction: Plan a readv via the index, read the deltas in forward IO, apply each delta. Total IO: sum(deltas) + deltacount*index overhead. Fetch/stacked reconstruction: Plan a readv via the index, using local basis texts where possible. Then readv locally and remote and apply deltas. Total IO as for solo reconstruction.

Things to keep

Reasonable sizes ‘amount read’ from remote machines to reconstruct an arbitrary text: Reading 5MB for a 100K plain text is not a good trade off. Reading (say) 500K is probably acceptable. Reading ~100K is ideal. However, it’s likely that some texts (e.g NEWS versions) can be stored for nearly-no space at all if we are willing to have unbounded IO. Profiling to set a good heuristic will be important. Also allowing users to choose to optimise for a server environment may make sense: paying more local IO for less compact storage may be useful.

Things to remove

Index scatter gather IO. Doing hundreds or thousands of index lookups is very expensive, and doing that per file just adds insult to injury.

Partioned compression amongst files.

Scatter gather IO when reconstructing texts: linear forward IO is better.


Merges combine texts from multiple versions to create a new version. Deltas add new text to existing files and remove some text from the same. Getting high compression means reading some base and then a chain of deltas (could be a tree) to gain access to the thing that the final delta was made against, and that delta. Rather than composing all these deltas, we can just just perform the final diff against the base text and the serialised invidual deltas. If the diff algorithm can reuse out of order lines from previous texts (e.g. storing AB -> BA as pointers rather than delete and add, then the presence of any previously stored line in a single chain can be reused. One such diff algorithm is xdelta, another reasonable one to consider is plain old zlib or lzma. We could also use bzip2. One advantage of using a generic compression engine is less python code. One advantage of preprocessing line based deltas is that we reduce the window size for the text repeated within lines, and that will help compression by a simple entropy compressor as a post processor. lzma appears fantastic at compression - 420MB of NEWS files down to 200KB. so window size appears to be a key determiner for efficiency.

Delta strategy

Very big objects - no delta. I plan to kick this in at 5MB initially, but once the codebase is up and running, we can tweak this to

Very small objects - no delta? If they are combined with a larger zlib object why not? (Answer: because zlib’s window is really small)

Other objects - group by fileid (gives related texts a chance, though using a file name would be better long term as e.g. COPYING and COPYING from different projects could combine). Then by reverse topological graph(as this places more recent texts at the front of a chain). Alternatively, group by size, though that should not matter with a large enough window. Finally, delta the texts against the current output of the compressor. This is essentially a somewhat typed form of sliding window dictionary compression. An alternative implementation would be to just use zlib, or lzma, or bzip2 directory.

Unfortunately, just using entropy compression forces a lot of data to be output by the decompressor - e.g. 420MB in the NEWS sample corpus. When we only want a single 55K text thats inefficient. (An initial test took several seconds with lzma.)

The fastest to implement approach is probably just ‘diff output to date and add to entropy compressor’. This should produce reasonable results. As delta chain length is not a concern (only one delta to apply ever), we can simply cap the chain when the total read size becomes unreasonable. Given older texts are smaller we probably want some weighted factor of plaintext size.

In this approach, a single entropy compressed region is read as a unit, giving the lower bound for IO (and how much to read is an open question - what byte offset of compressed data is sufficient to ensue that the delta-stream contents we need are reconstructable. Flushing, while possible, degrades compression(and adds overhead - we’d be paying 4 bytes per record guaranteed). Again - tests will be needed.

A nice possibility is to output mpdiff compatible records, which might enable some code reuse. This is more work than just diff (current_out, new_text), so can wait for the concept to be proven.

Implementation Strategy

Bring up a VersionedFiles object that implements this, then stuff it into a repository format. zlib as a starting compressor, though bzip2 will probably do a good job.