Reduce memory allocations in index #7138
Conversation
| @@ -515,6 +523,16 @@ func (d *DatabaseIndex) DropSeries(keys []string) { | |||
| atomic.AddInt64(&d.stats.NumSeries, -nDeleted) | |||
| } | |||
|
|
|||
| // Dereference removes all references to data within b and moves them to the heap. | |||
| func (d *DatabaseIndex) Dereference(b []byte) { | |||
| d.mu.Lock() | |||
There was a problem hiding this comment.
Probably needs to be an RLock so that a slow delete doesn't lock up the whole database.
There was a problem hiding this comment.
Dereference changes the underlying series so I think we need to write lock. I'm not sure how to get around it unless we get a list of series ids and then iterate over it in chunks.
There was a problem hiding this comment.
This is taking a lock on the index though, not the series. A write lock on the series makes sense though.
There was a problem hiding this comment.
Fixed in f6a1b2a.
benbjohnson
force-pushed
the
optimize-shard-open
branch
5 times, most recently
from
3acdf01 to
f6a1b2a
Compare
| @@ -1118,6 +1130,11 @@ func (p *purger) purge() { | |||
| p.mu.Lock() | |||
| for k, v := range p.files { | |||
| if !v.InUse() { | |||
| // Remove any mmap references held by the index. | |||
| if p.fileStore.dereferencer != nil { | |||
| v.deref(p.fileStore.dereferencer) | |||
There was a problem hiding this comment.
This also needs to be called within FileStore.Replace and FileStore.Remove. The Purger just handles removing files that are still in use by a query after a compaction has run.
There was a problem hiding this comment.
I added it to FileStore.Replace in 6614e0b but FileStore.Remove doesn't seem to actually close the file so it doesn't seem like the right place for dereference. Is it the purger that actually closes those files?
benbjohnson
force-pushed
the
optimize-shard-open
branch
from
f6a1b2a to
6614e0b
Compare
|
Tested this quite a bit yesterday and memory usage after restarts drops about ~30-40%. Objects in the heap was also reduced ~5-6x. Startup times are also reduced quite a bit. There is not much of a change if there is no restart. If we can figure out a solution for @joelegasse concerns then this should be good to merge. |
This commit changes the index to point to index data in the shards instead of keeping it in-memory on the heap.
benbjohnson
force-pushed
the
optimize-shard-open
branch
from
6614e0b to
8aa224b
Compare
|
👍 |
|
cc @daviesalex @sebito91 This will be in tonights nightly. |
Updates to interface with influxdb model. See: influxdata/influxdb#7138
Updates to interface with influxdb model. See: influxdata/influxdb#7138
Overview
This commit changes the index to point to index data in the shards instead of keeping it in-memory on the heap.
NOTE: This pull request is still missing dereferencing from the mmap when a shard is closed.
/cc @pauldix @jwilder
TODO