The improvements in the range-export code lead to avoid reading most blocks
twice, as well as to allowing some blocks to be written to disk multiple times.
The cache hit-rate went down from being close to 50% to a maximum of 12% at
the very end of the export. The reason is that most CIDs are never read twice
since they are correctly tracked in the CID set.
These numbers do not support the maintenance of the CachingBlockstore
code. Additional testing shows that removing it has similar memory-usage
behaviour and about 5 minute-faster execution (around 10%).
Less code to maintain and less options to mess up with.
This commit moderately refactors the ranged export code. It addresses several
problems:
* Code does not finish cleanly and things hang on ctrl-c
* Same block is read multiple times in a row (artificially increasing cached
blockstore metrics to 50%)
* It is unclear whether there are additional races (a single worker quits
when reaching height 0)
* CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or
so). Some blocks appear up to 5 times.
* Using pointers for tasks where it is not necessary.
The changes:
* Use a FIFO instead of stack: simpler implementation as its own type. This
has not proven to be much more memory-friendly, but it has not made things
worse either.
* We avoid a probably not small amount of allocations by not using
unnecessary pointers.
* Fix duplicated blocks by atomically checking+adding to CID set.
* Context-termination now works correctly. Worker lifetime is correctly tracked and all channels
are closed, avoiding any memory leaks and deadlocks.
* We ensure all work is finished before finishing, something that might have
been broken in some edge cases previously. In practice, we would not have
seen this except perhaps in very early snapshots close to genesis.
Initial testing shows the code is currently about 5% faster. Resulting
snapshots do not have duplicates so they are a bit smaller. We have manually
verified that no CID is lost versus previous results, with both old and recent
snapshots.
This first commit contains the first and second implementation stabs (after
primary review by @hsanjuan), using a stack for task buffering.
Known issues: ctrl-c (context cancellation) results in the export code getting
deadlocked. Duplicate blocks in exports. Duplicate block reads from store.
Original commit messages:
works
works against mainnet and calibnet
feat: add internal export api method
- will hopfully make things faster by not streaming the export over the json rpc api
polish: better file nameing
fix: potential race in marking cids as seen
chore: improve logging
feat: front export with cache
fix: give hector a good channel buffer on this shit
docsgen
We need this for NV16 to include code in chain snapshots.
NOTE: I've also checked the splitstore, and we appear to be doing the
right thing there already.
