add objects list caching for boltdb-shipper index store to reduce object storage list api calls #5160
Conversation
sandeepsukhani
force-pushed
the
boltdb-shipper-list-caching
branch
from
1d20707 to
5bd0127
Compare
…ect storage list api calls
sandeepsukhani
force-pushed
the
boltdb-shipper-list-caching
branch
from
5bd0127 to
9d2a103
Compare
| return nil, nil, fmt.Errorf("invalid prefix %s", prefix) | ||
| } | ||
|
|
||
| if !c.cacheBuiltAt.Add(cacheTimeout).After(time.Now()) { |
There was a problem hiding this comment.
Could also be written as:
| if !c.cacheBuiltAt.Add(cacheTimeout).After(time.Now()) { | |
| if time.Since(c.cacheBuiltAt) > cacheTimeout { |
which I find easier to read.
| select { | ||
| case c.rebuildCacheChan <- struct{}{}: | ||
| c.err = nil | ||
| c.err = c.buildCache(ctx) | ||
| <-c.rebuildCacheChan | ||
| if c.err != nil { | ||
| level.Error(util_log.Logger).Log("msg", "failed to build cache", "err", c.err) | ||
| } | ||
| default: | ||
| for !c.cacheBuiltAt.Add(cacheTimeout).After(time.Now()) && c.err == nil { | ||
| time.Sleep(time.Millisecond) | ||
| } | ||
| } |
There was a problem hiding this comment.
I have a hard time understanding why you chose to use a channel here. I assume to block concurrent access on List(). First call is building the cache while all others wait until cache is built?
There was a problem hiding this comment.
Yeah, just first or one of the concurrent calls to list should get to build the cache while others wait for it to finish successfully or with error. I will add a comment to make it clearer.
| c.tablesMtx.Lock() | ||
| defer c.tablesMtx.Unlock() | ||
|
|
||
| c.tables = map[string]*table{} |
There was a problem hiding this comment.
Could we decrease the lock time by assigning c.tables at the very end?
| c.tablesMtx.Lock() | |
| defer c.tablesMtx.Unlock() | |
| c.tables = map[string]*table{} | |
| new_tables := map[string]*table{} | |
| ... | |
| c.tablesMtx.Lock() | |
| defer c.tablesMtx.Unlock() | |
| c.tables = new_tables | |
| c.cacheBuiltAt = time.Now() | |
| return nil |
There was a problem hiding this comment.
We want to keep it locked until we build the cache to avoid returning stale results. Most of these list calls happen async so I am refreshing the cache on demand instead of running a goroutine refreshing it every min since we usually do these operations every 5 mins in index-gateway and 10 mins in compactor by default.
| objects, commonPrefixes, err := cachedObjectClient.List(context.Background(), "", "") | ||
| require.NoError(t, err) | ||
| require.Equal(t, 1, objectClient.listCallsCount) | ||
| require.Equal(t, objects, []chunk.StorageObject{}) |
There was a problem hiding this comment.
Arguments of the Equal function are in "incorrect" order:
| require.Equal(t, objects, []chunk.StorageObject{}) | |
| require.Equal(t, []chunk.StorageObject{}, objects) |
The function interface is
func Equal(t TestingT, expected interface{}, actual interface{}, msgAndArgs ...interface{})
This isn't a problem as long as expected and actual are equal, but the test error message is misleading in case they aren't.
There was a problem hiding this comment.
Guess this is not only a problem in your test, but we have that all over the place.
There was a problem hiding this comment.
yeah, sorry I messed up the order. Fixed it.
| } | ||
| default: | ||
| for time.Since(c.cacheBuiltAt) >= cacheTimeout && c.err == nil { | ||
| time.Sleep(time.Millisecond) |
There was a problem hiding this comment.
For loop and time.Sleep is a no no !
You want to use the promise pattern instead. Not sure if we can avoid a lock/ RW lock.
There was a problem hiding this comment.
I added a sync.WaitGroup to make all the goroutines attempting to build the cache to wait until the operation gets over. Can you please check now whether it looks good?
|
Fixes #5018 |
What this PR does / why we need it:
We, as of now, do a LIST calls per table when we need to find its objects.
If someone has a lot of tables cached locally or has query readiness set to a large number of days, it results in many list calls because each querier tries to sync tables every 5 mins by default.
This PR reduces the number of LIST calls we make when using hosted object stores(S3, GCS, Azure Blob Storage and Swift) as a shared store for boltdb-shipper.
The idea is to do a flat listing of objects supported by hosted object stores mentioned above and cache it until it goes stale.
Special notes for your reviewer:
Since caching requires a flat listing supported only by hosted object stores, I have added a
prefixedObjectClient, making the implementation somewhat cleaner.prefixedObjectClienttakes care of adding/removing configured object prefix to the keys. WithoutprefixedObjectClient, we will have to also make the caching client aware of object prefixes.Checklist