perf(core): Fix performance issue in type filter #9065
Conversation
harshil-goel
requested review from
mangalaman93,
meghalims,
billprovince and
joshua-goldstein
as code owners
dgraph-bot
added
area/core
harshil-goel
force-pushed
the
harshil-goel/fix-type-perf
branch
2 times, most recently
from
b31cea7
to
8375b71
Compare
posting/list.go
Outdated
| numNormalPostingsRead := 0 | ||
| defer func() { | ||
| if numNormalPostingsRead < numDeletePostingsRead { | ||
| glog.V(3).Infof("During iterate on posting list, we read %d set postings, %d delete postings"+ |
There was a problem hiding this comment.
Can we clarify this message in some way to be more useful to someone who is not familiar with internal workings? Also include which posting list if that is available.
I think this means that the badger values representing a posting list (in 256KB chunks IIRC) contained many deleted structures, and are doing more than 50% of the data movement for non-useful deleted data.
If so something like: "High proportion of deleted data observed for posting list {l.key}: total = {numNormal+numdeleted}, percent deleted = {numDel / (numnormal+numdel) * 100}%".
| @@ -48,7 +48,7 @@ const ( | |||
| `client_key=; sasl-mechanism=PLAIN; tls=false;` | |||
| LimitDefaults = `mutations=allow; query-edge=1000000; normalize-node=10000; ` + | |||
| `mutations-nquad=1000000; disallow-drop=false; query-timeout=0ms; txn-abort-after=5m; ` + | |||
| ` max-retries=10;max-pending-queries=10000;shared-instance=false` | |||
| ` max-retries=10;max-pending-queries=10000;shared-instance=false;type-filter-uid-limit=10` | |||
There was a problem hiding this comment.
Thinking about this... If a customer is large enough to have performance concerns, my guess is the dgraph.type index is going to be very large (or include some very large index entries that drive the performance profile). In that case, perhaps we should optimize for this case, and consider what uid-limit will balance performance for 1M or more UIDs? I suspect that will be more like 100 or more.
Maybe not something to delay/retest for, but consider.
There was a problem hiding this comment.
Agreed. The number 10 is too small for such optimization. But we can defer this until we see another case and get some validation of type-filter-uid-limit.
Currently when we do queries like `func(uid: 0x1) @filter(type)`. We retrieve the entire type index. Sometimes, when the index is too big, fetching the index is quite slow. We realised that if we know we only want to check few `uids` are of the same, then we can just check those `uids` directly. Right now we are hard coding the number of `uids` threshold. This could be improved with a more statistical based model, where we figure out how many items does the type index have, how many we need to check.
Currently when we do queries like `func(uid: 0x1) @filter(type)`. We retrieve the entire type index. Sometimes, when the index is too big, fetching the index is quite slow. We realised that if we know we only want to check few `uids` are of the same, then we can just check those `uids` directly. Right now we are hard coding the number of `uids` threshold. This could be improved with a more statistical based model, where we figure out how many items does the type index have, how many we need to check.
Currently when we do queries like
func(uid: 0x1) @filter(type). We retrieve the entire type index. Sometimes, when the index is too big, fetching the index is quite slow. We realised that if we know we only want to check fewuidsare of the same, then we can just check thoseuidsdirectly. Right now we are hard coding the number ofuidsthreshold. This could be improved with a more statistical based model, where we figure out how many items does the type index have, how many we need to check.