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.
This makes Go 1.15 test/vet happy, avoiding "conversion from untyped int
to string yields a string of one rune" warning where we do
string(KeyTypeWhatever) in namespaced.go.
It also clarifies and enforces the currently allowed range of these
numbers so I think it's fine.
This matches the convention of the stdlib and avoids ambiguity: when
customErr{} and &customErr{} both implement error, client code needs to
check for both.
Memory use should remain the same, since storing a non-pointer type in
an interface value still copies the value to the heap.
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 reduces the size of our write batches before we flush them. This
has two effects: reducing the amount of data lost if we crash when
updating the database, and reducing the amount of memory used when we do
large updates without checkpoint (e.g., deleting a folder).
I ran our SyncManyFiles benchmark as it is the one doing most
transactions, however there was no relevant change in any metric (it's
limited by our fsync I expect). This is good as any visible change would
just be a decrease in performance.
I don't have a benchmark on deleting a large folder, taking that part on
trust for now...
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...
- In the few places where we wrap errors, use the new Go 1.13 "%w"
construction instead of %s or %v.
- Where we create errors with constant strings, consistently use
errors.New and not fmt.Errorf.
- Remove capitalization from errors in the few places where we had that.
If we decide to recalculate the metadata we shouldn't start from
whatever we loaded from the database, as that data is wrong. We should
start from a clean slate.
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.
lib/db: Recover sequence number and metadata on startup (fixes#6335)
If we crashed after writing new file entries but before updating
metadata in the database the sequence number and metadata will be wrong.
This fixes that.
We could potentially get a snapshot and then fail to get a releaser,
leaking the snapshot. This takes the releaser first and makes sure to
release it on snapshot error.
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
* lib/db, lib/protocol: Compact FileInfo and BlockInfo alignment
This fixes the following two lint warnings
FileInfo: struct of size 160 bytes could be of size 136 bytes
BlockInfo: struct of size 48 bytes could be of size 40 bytes
by reordering fields in alignment order (64 bit fields, then 32 bit
fields, then 16 bit fields (if any), then small ones). The end result is
a slightly less aesthetically pleasing struct field order, but since
these are the objects we often juggle in bulk and keep large queues of I
think it's worth it.
It's a micro optimization, but a cheap one.
As foretold by the prophecy, "once the database refactor is merged, then
shall appear a request to propagate errors from the store known
throughout the land as the NamedspacedKV, and it shall be good".
This PR does two things, because one lead to the other:
- Move the leveldb specific stuff into a small "backend" package that
defines a backend interface and the leveldb implementation. This allows,
potentially, in the future, switching the db implementation so another
KV store should we wish to do so.
- Add proper error handling all along the way. The db and backend
packages are now errcheck clean. However, I drew the line at modifying
the FileSet API in order to keep this manageable and not continue
refactoring all of the rest of Syncthing. As such, the FileSet methods
still panic on database errors, except for the "database is closed"
error which is instead handled by silently returning as quickly as
possible, with the assumption that we're anyway "on the way out".
This is the result of:
- Changing build.go to take the protobuf version from the modules
instead of hardcoded
- `go get github.com/gogo/protobuf@v1.3.0` to upgrade
- `go run build.go proto` to regenerate our code
This introduces a better set of defaults for large databases. I've
experimentally determined that it results in much better throughput in a
couple of scenarios with large databases, but I can't give any
guarantees the values are always optimal. They're probably no worse than
the defaults though.
This adds a set of magical environment variables that can be used to
tweak the database parameters. It's totally undocumented and not
intended to be a long term or supported thing.
It's ugly, but there is a backstory. I have a couple of large
installations where the database options are inefficient or otherwise
suboptimal (24/7 compaction going on and stuff like that). I don't
*know* the correct database parameters, nor yet the formula or method to
derive them by, so this requires experimentation. Experimentation needs
to happen partly in production, and rolling out new builds for every
tweak isn't practical. This provides override points for all reasonable
values, while not changing anything by default.
Ideally, at the end of such experimentation, we'll know which values are
relevant to change and in what manner, and can provide a more user
friendly knob to do so - or do it automatically based on the database
size.
Flush the batch when exceeding a certain size, instead of when reaching a number
of batched operations.
Move batch to lowlevel to be able to use it in NamespacedKV.
Increase the leveldb memory buffer from 4 to 16 MiB.
To do so the BlockMap struct has been removed. It behaves like any other prefixed
part of the database, but was not integrated in the recent keyer refactor. Now
the database is only flushed when files are in a consistent state.
There was a problem in iterating the sequence index that could result
in missing updates. The issue is that while the index was (correctly)
iterated in a snapshot, the actual file infos were read dirty outside of
the snapshot. This fixes this by doing the reads inside the snapshot,
and also updates a couple of other places that did the same thing more
or less harmfully (I didn't investigate).
To avoid similar issues in the future I did some renaming of the
getFile* methods - the ones in a transaction are just getFile, while the
ones directly on the database are variants of getFileDirty to highlight
what's going on.
* go mod init; rm -rf vendor
* tweak proto files and generation
* go mod vendor
* clean up build.go
* protobuf literals in tests
* downgrade gogo/protobuf
This adds a thin type that holds the state associated with the
leveldb.DB, leaving the huge Instance type more or less stateless. Also
moves some keying stuff into the DB package so that other packages need
not know the keying specifics.
(This does not, yet, fix the cmd/stindex program, in order to keep the
diff size down. Hence the keying constants are still exported.)
The problem here is that we would update the sequence index before
updating the FileInfos, which would result in a high sequence number
pointing to a low-sequence FileInfo. The index sender would pick up the
high sequence number, send the old file, and think everything was good.
On the receiving side the old file is a no-op and ignored. The file
remains out of sync until another update for it happens.
This fixes that by correcting the order of operations in the database
update: first we remove old sequence index entries, then we update the
FileInfos (which now don't have anything pointing to them) and then we
add the sequence indexes (which the index sender can see).
The other option is to add "proper" transactions where required at the
database layer. I actually have a branch for that, but it's literally
thousands of lines of diff and I'm putting that off for another day as
this solves the problem...
Adds a receive only folder type that does not send changes, and where the user can optionally revert local changes. Also changes some of the icons to make the three folder types distinguishable.
We have the invalid bit to indicate that a file isn't good. That's enough for remote devices. For ourselves, it would be good to know sometimes why the file isn't good - because it's an unsupported type, because it matches an ignore pattern, or because we detected the data is bad and we need to rescan it.
Or, and this is the main future reason for the PR, because it's a change detected on a receive only device. We will want something like the invalid flag for those changes, but marking them as invalid today means the scanner will rehash them. Hence something more fine grained is required.
This introduces a LocalFlags fields to the FileInfo where we can stash things that we care about locally. For example,
FlagLocalUnsupported = 1 << 0 // The kind is unsupported, e.g. symlinks on Windows
FlagLocalIgnored = 1 << 1 // Matches local ignore patterns
FlagLocalMustRescan = 1 << 2 // Doesn't match content on disk, must be rechecked fully
The LocalFlags fields isn't sent over the wire; instead the Invalid attribute is calculated based on the flags at index sending time. It's on the FileInfo anyway because that's what we serialize to database etc.
The actual Invalid flag should after this just be considered when building the global state and figuring out availability for remote devices. It is not used for local file index entries.
To optimize WithNeed, which is called for the local device whenever an index
update is received. No tracking for remote devices to conserve db space, as
WithNeed is only queried for completion.
This adds a couple of utilities for transporting databases in JSON and a
test to load a database and verify a couple of invalid bits. The test
itself is quite pointless at the moment, but it lays the groundwork for
testing the migration of this data in the next step (after the invalid
bit should be changed to local flags for local files).
When that happens we need to have a database in the old format already
there in order to be able to test the migration.
Instead of walking and unmarshalling the entire db and sorting the resulting
file infos by sequence, add store device keys by sequence number in the
database. Thus only the required file infos need be unmarshalled and are already
sorted by index.
When dropping delta index IDs due to upgrade, only drop our local one.
Previously, when dropping all of them, we would trigger a full send in
both directions on first connect after upgrade. Then the other side
would upgrade, doing the same thing. Net effect is full index data gets
sent twice in both directions.
With this change we just drop our local ID, meaning we will send our
full index on first connect after upgrade. When the other side upgrades,
they will do the same. This is a bit less cruel.
Unignored files are marked as conflicting while scanning, which is then resolved
in the subsequent pull. Automatically reconciles needed items on send-only
folders, if they do not actually differ except for internal metadata.
* lib/db: Don't panic on negative counts (fixes#4659)
So, negative counts should never happen and hence the original idea to
panic. However, this sucks as the panic will happen in a folder runner,
be automatically swallowed by suture, and the runner gets restarted but
now we are in a bad state. (Related: #4758)
At the time of writing the global list is somewhat in flux (we've
changed how ignored files are handled, invalid bits, etc.) and I think
that can cause unusual conditions here. Hence just fixing up the numbers
instead until the next full recount.
When scanner.Walk detects a change, it now returns the new file info as well as the old file info. It also finds deleted and ignored files while scanning.
Also directory deletions are now always committed to db after their children to prevent temporary failure on remote due to non-empty directory.
