Add eviction cycle metrics #4144
Conversation
src/bucket/BucketManagerImpl.cpp
Outdated
| mBytesScannedForEviction, | ||
| mIncompleteBucketScans); | ||
| mBucketList->scanForEviction( | ||
| mApp, ltx, ledgerSeq, mEntriesEvicted, mBytesScannedForEviction, |
There was a problem hiding this comment.
Why not make these metrics live in Bucketlist instead of passing them here? They don't seem to be used outside of Bucketlist anyway
There was a problem hiding this comment.
This is the standard interface we use for all BucketList related metrics. The BucketList itself is a very light weight class (the only member is std::vector<BucketLevel> mLevels;). We could do this, but we would also have to change the BucketList constructor from BucketList() to BucketList(Application), which would be annoying for tests and a bit of a refactor.
There was a problem hiding this comment.
Ok, I flagged this because this function signature is getting a bit out of hand. Could we use a struct here?
| if (metrics.has_value()) | ||
| { | ||
| ++metrics->numEntriesEvicted; | ||
| metrics->evictedEntriesAgeSum += |
There was a problem hiding this comment.
why do we need to sum it up like this? we could use a histrogram (that would live inside Bucket), and update it every time we evict a single entry instead of manually aggregating evictions.
There was a problem hiding this comment.
We actually don't want a histogram and specifically want the per-cycle aggregation. A histogram for this metric is misleading. Given the nature of the eviction scan, entries evicted in level 8 will almost certainly be "older" than entries evicted in level 6. Within a given cycle, the number of evicted entries is very noisy from level to level. We want to abstract away that noise. We only care if the average age increases cycle-to-cycle. This would be pretty involved to graph in Grafana if we use a histogram, as the cycle length itself is variable. This seemed significantly easier, as it logs the only metric we actually care about (average cycle-to-cycle) in a way that is easy to analyze in Grafana.
src/bucket/BucketList.h
Outdated
| { | ||
| // Evicted entry "age" is the delta between its liveUntilLedger and the | ||
| // ledger when the entry is actually evicted | ||
| uint64_t evictedEntriesAgeSum; |
There was a problem hiding this comment.
uninitialized variables
There was a problem hiding this comment.
(also see my other comment regarding average age, I think you can remove this struct completely, and only keep evictionCycleStartLedger)
SirTyson
force-pushed
the
eviction-metrics
branch
from
ccbc81e
to
7d6b26b
Compare
SirTyson
force-pushed
the
eviction-metrics
branch
from
7d6b26b
to
407d836
Compare
|
r+ 407d836 |
Description
Resolves #4137
Adds additional metrics for eviction scans. Specifically, records the period of eviction scans in ledgers and the average amount of time it takes for an entry to get evicted after expiring.
Checklist
clang-formatv8.0.0 (viamake formator the Visual Studio extension)