The seen BlobSet always contained a subset of the entries in blobs.
Thus use blobs instead and avoid the memory overhead of the second set.
Suggested-by: Alexander Weiss <alex@weissfam.de>
- Add Open() functionality to dir
- only access index for blobs when file is read
- Implement NodeOpener and put one-time file stuff there
- Add comment about locking as suggested by bazil.org/fuse
=> Thanks at Michael Eischer for suggesting the last two improvements
Calling `Close()` on the rclone backend sometimes failed during test
execution with 'signal: Broken pipe'. The stdio connection closed both
the stdin and stdout file descriptors at the same moment, therefore
giving rclone no chance to properly send any final http2 data frames.
Now the stdin connection to rclone is closed first and will only be
forcefully closed after a timeout. In case rclone exits before the
timeout then the stdio connection will be closed normally.
restic did not notice when the rclone subprocess exited unexpectedly.
restic manually created pipes for stdin and stdout and used these for the
connection to the rclone subprocess. The process creating a pipe gets
file descriptors for the sender and receiver side of a pipe and passes
them on to the subprocess. The expected behavior would be that reads or
writes in the parent process fail / return once the child process dies
as a pipe would now just have a reader or writer but not both.
However, this never happened as restic kept the reader and writer
file descriptors of the pipes. `cmd.StdinPipe` and `cmd.StdoutPipe`
close the subprocess side of pipes once the child process was started
and close the parent process side of pipes once wait has finished. We
can't use these functions as we need access to the raw `os.File` so just
replicate that behavior.
The test now uses the fact that the sort is stable. It's not guaranteed
to be, but the test is cleaner and more exhaustive. sortCachedPacksFirst
no longer needs a return value.
If a data blob and a tree blob with the same ID (= same content) exist,
then the checker did not report a data or tree blob as unused when the
blob of the other type was still in use.
The `DuplicateTree` flag is necessary to ensure that failures cannot be
swallowed. The old checker implementation ignores errors from LoadTree
if the corresponding tree was already checked.
Backups traverse the file tree in depth-first order and saves trees on
the way back up. This results in tree packs filled in a way comparable
to the reverse Polish notation. In order to check tree blobs in that
order, the treeFilter would have to delay the forwarding of tree nodes
until all children of it are processed which would complicate the
implementation.
Therefore do the next similar thing and traverse the tree in depth-first
order, but process trees already on the way down. The tree blob ids are
added in reverse order to the backlog, which is once again reverted when
removing the ids from the back of the backlog.
The blobRefs map and the processedTrees IDSet are merged to reduce the
memory usage. The blobRefs map now uses separate flags to track blob
usage as data or tree blob. This prevents skipping of trees whose
content is identical to an already processed data blob. A third flag
tracks whether a blob exists or not, which removes the need for the
blobs IDSet.
Even though the checkTreeWorker skips already processed chunks,
filterTrees did queue the same tree blob on every occurence. This
becomes a serious performance bottleneck for larger number of snapshots
that cover mostly the same directories. Therefore decode a tree blob
exactly once.
The benchmark was actually testing the speed of index lookups.
name old time/op new time/op delta
SaveAndEncrypt-8 101ns ± 2% 31505824ns ± 1% +31311591.31% (p=0.000 n=10+10)
name old speed new speed delta
SaveAndEncrypt-8 41.7TB/s ± 2% 0.0TB/s ± 1% -100.00% (p=0.000 n=10+10)
name old alloc/op new alloc/op delta
SaveAndEncrypt-8 1.00B ± 0% 20989508.40B ± 0% +2098950740.00% (p=0.000 n=10+10)
name old allocs/op new allocs/op delta
SaveAndEncrypt-8 0.00 123.00 ± 0% +Inf% (p=0.000 n=10+9)
(The actual speed is ca. 131MiB/s.)
A side remark to the definition of Index.blob:
Another possibility would have been to use:
blob map[restic.BlobHandle]*indexEntry
This would have led to the following sizes:
key: 32 + 1 = 33 bytes
value: 8 bytes
indexEntry: 8 + 4 + 4 = 16 bytes
each packID: 32 bytes
To save N index entries, we would therefore have needed:
N * OF * (33 + 8) bytes + N * 16 + N * 32 bytes / BP = N * 82 bytes
More precicely, using a pointer instead of a direct entry is the better memory choice if:
OF * 8 bytes + entrysize < OF * entrysize <=> entrysize > 8 bytes * OF/(OF-1)
Under the assumption of OF=1.5, this means using pointers would have been the better choice
if sizeof(indexEntry) > 24 bytes.
