[multiverse RPC]: add better universe root node cache #1169
Conversation
Pull Request Test Coverage Report for Build 11920129966Details
💛 - Coveralls |
| // multiverse cache. | ||
| func DefaultMultiverseCacheConfig() MultiverseCacheConfig { | ||
| return MultiverseCacheConfig{ | ||
| ProofsPerUniverse: 5, |
There was a problem hiding this comment.
Temp for this commit?
We should inherit the defaults defined in the existing package here.
There was a problem hiding this comment.
I think the original value for this was way too high. Because this is the number of proofs we cache per universe (and we don't have a limit on the number of universes we cache proofs for, it's a lnutils.SyncMap here and not an lru.Cache), we should be quite conservative here.
Perhaps we should turn this into an lru.Cache too, same as with the universeLeafPageCache?
There was a problem hiding this comment.
Ah ok, the original intent was for that to be a instance wide value (eg: only cache 50k proofs total).
I think before merge, we should deploy a version of this on testnet/mainnet to observe the cache hit/miss rate.
guggero
force-pushed
the
universe-cache
branch
2 times, most recently
from
8c63baa to
b63f383
Compare
|
Addressed all comments and added proper unit tests. Good for full review now. |
guggero
force-pushed
the
universe-cache
branch
from
b63f383 to
b04ddf1
Compare
|
I added a test that calculates the approximate size of the entries in the cache: With around 170k roots currently in existence on mainnet, that means the new cache would only use ~34 MBytes of memory. |
| params := sqlc.UniverseRootsParams{ | ||
| SortDirection: sqlInt16(universe.SortAscending), | ||
| NumOffset: 0, | ||
| NumLimit: universe.MaxPageSize, |
There was a problem hiding this comment.
In cases like this, I wonder if it's better to just read the entire set of roots in one go. So one single query, instead of a bunch of smaller ones. With the current amount of assets on mainnet, we'll end up doing 300 or so queries here to read them all out. Some benchmarking would likely help us find the sweet spot here re page size to read.
Non-blocking.
| @@ -271,6 +475,7 @@ func (r *rootNodeCache) cacheRoots(q universe.RootNodesQuery, | |||
|
|
|||
| rootIndex := r.rootIndex.Load() | |||
| for _, rootNode := range rootNodes { | |||
| rootNode := rootNode | |||
There was a problem hiding this comment.
FWIW, with the latest version of Go, we should no longer need to do this reassignment trick.
| // multiverse cache. | ||
| func DefaultMultiverseCacheConfig() MultiverseCacheConfig { | ||
| return MultiverseCacheConfig{ | ||
| ProofsPerUniverse: 5, |
There was a problem hiding this comment.
Ah ok, the original intent was for that to be a instance wide value (eg: only cache 50k proofs total).
I think before merge, we should deploy a version of this on testnet/mainnet to observe the cache hit/miss rate.
This commit adds a completely new cache that is tailor made for the queries the syncer makes (ascending, without any amounts by ID). This additional cache has all roots (without details), so the existing cache will only be needed for custom queries that aren't coming from the syncer. We should be able to down-size the existing cache quite a bit, as it should only serve manual requests or requests coming from an UI. We'll make the size configurable in an upcoming commit.
This commit adds a test that calculates the approximate memory
consumption of a full syncer cache.
The results look pretty good:
=== RUN TestSyncerCacheMemoryUsage
multiverse_cache_test.go:214: Generated 500 roots in 26.583318ms
=== RUN TestSyncerCacheMemoryUsage/500_roots
multiverse_cache_test.go:224: Memory usage for 500 roots: 139496 bytes
multiverse_cache_test.go:226: Memory usage per root: 278 bytes
multiverse_cache_test.go:228: Benchmark took 8.836µs
--- PASS: TestSyncerCacheMemoryUsage/500_roots (0.01s)
multiverse_cache_test.go:214: Generated 5000 roots in 257.823179ms
=== RUN TestSyncerCacheMemoryUsage/5000_roots
multiverse_cache_test.go:224: Memory usage for 5000 roots: 1073384 bytes
multiverse_cache_test.go:226: Memory usage per root: 214 bytes
multiverse_cache_test.go:228: Benchmark took 104.568µs
--- PASS: TestSyncerCacheMemoryUsage/5000_roots (0.01s)
multiverse_cache_test.go:214: Generated 170000 roots in 8.847171795s
=== RUN TestSyncerCacheMemoryUsage/170000_roots
multiverse_cache_test.go:224: Memory usage for 170000 roots: 34259176 bytes
multiverse_cache_test.go:226: Memory usage per root: 201 bytes
multiverse_cache_test.go:228: Benchmark took 1.820929ms
--- PASS: TestSyncerCacheMemoryUsage/170000_roots (0.08s)
--- PASS: TestSyncerCacheMemoryUsage (9.23s)
This commit allows for a much larger (32x) page size than before. This should allow us to fetch the complete set of roots even faster, given that we should have 100% cache hits for the syncer calls now. We also reduce the overall count of requests by 1 in almost all cases by detecting that the returned page isn't full anymore (instead of reading until we get an empty page).
guggero
force-pushed
the
universe-cache
branch
from
b04ddf1 to
e8065e4
Compare
Fixes #880.
Opening as draft, mainly looking for concept ACK.
@Roasbeef the idea here is the following:
sort=ascendingandWithAmountsById=false, which allows us to optimize the new cache and keep all root nodes in memory for those queriesWithAmountsById=true, which allows us to keep a much smaller cacheTODOs: