Improve MySQL restore performance #1131
Conversation
pkg/storages/storage/mergewriter.go
Outdated
| continue | ||
| } | ||
| rbytes := len(block) | ||
| wbytes, err := sink.Write(block) |
There was a problem hiding this comment.
io.Writer may write block partially without error
We should retry with the rest of the block until it's over or error returned.
There was a problem hiding this comment.
Acoarding to io.Writer documentation:
Write must return a non-nil error if it returns n < len(p).
ostinru
force-pushed
the
readahead
branch
4 times, most recently
from
2ea5738
to
f323485
Compare
| @@ -155,3 +196,19 @@ func (uploader *Uploader) UploadMultiple(objects []UploadObject) error { | |||
| } | |||
| return nil | |||
| } | |||
|
|
|||
| func (uploader *Uploader) ChangeDirectory(relativePath string) { | |||
There was a problem hiding this comment.
I guess that SetFolder(storage.Folder) would be more universal.
There was a problem hiding this comment.
SetFolder sounds like absolute path. Change Directory is like uniux cd
| ) | ||
|
|
||
| type BackupStreamMetadata struct { | ||
| Type string `json:"type"` |
There was a problem hiding this comment.
If this metadata does not consume much space, maybe just add it to the sentinel?
There was a problem hiding this comment.
it will increase coupling between uploader metadata and sentinel
internal/stream_metadata.go
Outdated
| } | ||
|
|
||
| func UploadBackupStreamMetadata(uploader UploaderProvider, metadata interface{}, backupName string) error { | ||
| sentinelName := MetadataNameFromBackup(backupName) |
There was a problem hiding this comment.
Well, this name interferes with the Postgres metadata.json file, but has the different purpose. Since this file contains only the stream-specific metadata maybe name it like stream_metadata.json?
| } | ||
|
|
||
| uploader := &SplitStreamUploader{ | ||
| Uploader: &Uploader{ |
There was a problem hiding this comment.
Just pass the Uploader to the NewSplitStreamUploader to avoid constructing the internal uploader. Actually, can you use the UploaderProvider interface here?
* Improve download speed by splitting backup stream in multiple blobs. Each of which can be decoded & decompressed independetly * use background readahead goroutine to separate decoding and decompressing stages. * add concurrent range-based S3-reader (WIP)
* save type of the backup in sentinel * use sentinel in backup-fetch command
* improve error handling
…allelism. However it is better to just increase `WALG_STREAM_SPLITTER_PARTITIONS`: in this case parallelism increased without adding unnecessary memory copy.
…se memory buffers
…hronously copy data with size greater than 2*blockSize.
* Hide all upload logic inside new SplitStreamUploader * use new Uploader in mongodb & mysql implementations
This patch improves wal-g-mysql backup-push/backup-fetch by splitting single-stream backup into multi-stream backups:
In order to provide database-agnostic approach we don't make assumptions about backup-stream wal-g getting from stdin.
Instead, wal-g splits backup stream in blocks of
WALG_STREAM_SPLITTER_BLOCK_SIZEbytes (instorages/splitmerge/splitreader.go). Then blocks are sent to one ofWALG_STREAM_SPLITTER_PARTITIONSoutput streams. Each stream compressed, encrypted and uploaded separately.Restore process is doing the opposite: it fetches multiple streams (utilize network by using multiple connections), then each stream can be decrypted and decompressed independently (utilize more CPU cores). Then it merges (in
storages/splitmerge/mergewriter.go) all streams into one stream.In my benchmarks
wal-g backup-fetch LATEST --turbo: