If the GC finds a key k that it wants to keep, it records that in a
Bloom filter. If a key k' can be removed but its hash collides with k,
it will be kept. Since the old Bloom filter code was completely
deterministic, the next run would encounter the same collision, assuming
k must still be kept.
A randomized hash function that uses all the SHA-256 bits solves this
problem: the second run has a non-zero probability of removing k', as
long as the Bloom filter is not completely full.
Group the global list of files by version, instead of having one flat list for all devices. This removes lots of duplicate protocol.Vectors.
Co-authored-by: Jakob Borg <jakob@kastelo.net>
This adds indirection of large version vectors in the same manner as we
already to block lists. The effect is the same: less duplicated data in
some situations.
To mitigate the impact for when this indirection
wouldn't be needed I've added an indirection cutoff for both blocks and
the new version vector stuff: we don't do the indirection at all for
small block lists or small version vectors, instead storing it directly
like we used to do. This is faster for small files and small setups.
This makes the GC runner a service that will stop fairly quickly when
told to.
As a bonus, STTRACE=app will print the service tree on the way out,
including any errors they've flagged.
As of the latest database checker we are again putting files without
blocks. I'm not 100% convinced that's a great idea, but we also do it
for ignored files apparently so it looks like we probably should support
it. This adds an escape hatch that must be manually enabled...
I was working on indirecting version vectors, and that resulted in some
refactoring and improving the existing block indirection stuff. We may
or may not end up doing the version vector indirection, but I think
these changes are reasonable anyhow and will simplify the diff
significantly if we do go there. The main points are:
- A bunch of renaming to make the indirection and GC not about "blocks"
but about "indirection".
- Adding a cutoff so that we don't actually indirect for small block
lists. This gets us better performance when handling small files as it
cuts out the indirection for quite small loss in space efficiency.
- Being paranoid and always recalculating the hash on put. This costs
some CPU, but the consequences if a buggy or malicious implementation
silently substituted the block list by lying about the hash would be bad.
This adds metadata updates to the same write batch as the underlying
file change. The odds of a metadata update going missing is greatly
reduced.
Bonus change: actually commit the transaction in recalcMeta.
The readWriteTransaction offered both commit() (the one to use) and
Commit() (via embedding) where the latter didn't close the read
transaction. This removes the lower cased variant in order to prevent
the mistake.
The only place where the mistake was made was the new gc runner, where
it would leave a read snapshot open forever.
Also retain the interval over restarts by storing last GC time in the
database. This to make sure that GC eventually happens even if the
interval is configured to a long time (say, a month).
* lib/db: Deduplicate block lists in database (fixes#5898)
This moves the block list in the database out from being just a field on
the FileInfo to being an object of its own. When putting a FileInfo we
marshal the block list separately and store it keyed by the sha256 of
the marshalled block list. When getting, if we are not doing a
"truncated" get, we do an extra read and unmarshal for the block list.
Old block lists are cleared out by a periodic GC sweep. The alternative
would be to use refcounting, but:
- There is a larger risk of getting that wrong and either dropping a
block list in error or keeping them around forever.
- It's tricky with our current database, as we don't have dirty reads.
This means that if we update two FileInfos with identical block lists in
the same transaction we can't just do read/modify/write for the ref
counters as we wouldn't see our own first update. See above about
tracking this and risks about getting it wrong.
GC uses a bloom filter for keys to avoid heavy RAM usage. GC can't run
concurrently with FileInfo updates so there is a new lock around those
operation at the lowlevel.
The end result is a much more compact database, especially for setups
with many peers where files get duplicated many times.
This is per-key-class stats for a large database I'm currently working
with, under the current schema:
```
0x00: 9138161 items, 870876 KB keys + 7397482 KB data, 95 B + 809 B avg, 1637651 B max
0x01: 185656 items, 10388 KB keys + 1790909 KB data, 55 B + 9646 B avg, 924525 B max
0x02: 916890 items, 84795 KB keys + 3667 KB data, 92 B + 4 B avg, 192 B max
0x03: 384 items, 27 KB keys + 5 KB data, 72 B + 15 B avg, 87 B max
0x04: 1109 items, 17 KB keys + 17 KB data, 15 B + 15 B avg, 69 B max
0x06: 383 items, 3 KB keys + 0 KB data, 9 B + 2 B avg, 18 B max
0x07: 510 items, 4 KB keys + 12 KB data, 9 B + 24 B avg, 41 B max
0x08: 1349 items, 12 KB keys + 10 KB data, 9 B + 8 B avg, 17 B max
0x09: 194 items, 0 KB keys + 123 KB data, 5 B + 634 B avg, 11484 B max
0x0a: 3 items, 0 KB keys + 0 KB data, 14 B + 7 B avg, 30 B max
0x0b: 181836 items, 2363 KB keys + 10694 KB data, 13 B + 58 B avg, 173 B max
Total 10426475 items, 968490 KB keys + 9202925 KB data.
```
Note 7.4 GB of data in class 00, total size 9.2 GB. After running the
migration we get this instead:
```
0x00: 9138161 items, 870876 KB keys + 2611392 KB data, 95 B + 285 B avg, 4788 B max
0x01: 185656 items, 10388 KB keys + 1790909 KB data, 55 B + 9646 B avg, 924525 B max
0x02: 916890 items, 84795 KB keys + 3667 KB data, 92 B + 4 B avg, 192 B max
0x03: 384 items, 27 KB keys + 5 KB data, 72 B + 15 B avg, 87 B max
0x04: 1109 items, 17 KB keys + 17 KB data, 15 B + 15 B avg, 69 B max
0x06: 383 items, 3 KB keys + 0 KB data, 9 B + 2 B avg, 18 B max
0x07: 510 items, 4 KB keys + 12 KB data, 9 B + 24 B avg, 41 B max
0x09: 194 items, 0 KB keys + 123 KB data, 5 B + 634 B avg, 11484 B max
0x0a: 3 items, 0 KB keys + 0 KB data, 14 B + 17 B avg, 51 B max
0x0b: 181836 items, 2363 KB keys + 10694 KB data, 13 B + 58 B avg, 173 B max
0x0d: 44282 items, 1461 KB keys + 61081 KB data, 33 B + 1379 B avg, 1637399 B max
Total 10469408 items, 969939 KB keys + 4477905 KB data.
```
Class 00 is now down to 2.6 GB, with just 61 MB added in class 0d.
There will be some additional reads in some cases which theoretically
hurts performance, but this will be more than compensated for by smaller
writes and better compaction.
On my own home setup which just has three devices and a handful of
folders the difference is smaller in absolute numbers of course, but
still less than half the old size:
```
0x00: 297122 items, 20894 KB keys + 306860 KB data, 70 B + 1032 B avg, 103237 B max
0x01: 115299 items, 7738 KB keys + 17542 KB data, 67 B + 152 B avg, 419 B max
0x02: 1430537 items, 121223 KB keys + 5722 KB data, 84 B + 4 B avg, 253 B max
...
Total 1947412 items, 151268 KB keys + 337485 KB data.
```
to:
```
0x00: 297122 items, 20894 KB keys + 37038 KB data, 70 B + 124 B avg, 520 B max
0x01: 115299 items, 7738 KB keys + 17542 KB data, 67 B + 152 B avg, 419 B max
0x02: 1430537 items, 121223 KB keys + 5722 KB data, 84 B + 4 B avg, 253 B max
...
0x0d: 18041 items, 595 KB keys + 71964 KB data, 33 B + 3988 B avg, 101109 B max
Total 1965447 items, 151863 KB keys + 139628 KB data.
```
* wip
* wip
* wip
* wip
Simon Frei
Audrius Butkevicius
greatroar
André Colomb
Jakob Borg