issue: #32995
To speed up the construction and querying of Bloom filters, we chose a
blocked Bloom filter instead of a basic Bloom filter implementation.
WARN: This PR is compatible with old version bf impl, but if fall back
to old milvus version, it may causes bloom filter deserialize failed.
In single Bloom filter test cases with a capacity of 1,000,000 and a
false positive rate (FPR) of 0.001, the blocked Bloom filter is 5 times
faster than the basic Bloom filter in both querying and construction, at
the cost of a 30% increase in memory usage.
- Block BF construct time {"time": "54.128131ms"}
- Block BF size {"size": 3021578}
- Block BF Test cost {"time": "55.407352ms"}
- Basic BF construct time {"time": "210.262183ms"}
- Basic BF size {"size": 2396308}
- Basic BF Test cost {"time": "192.596229ms"}
In multi Bloom filter test cases with a capacity of 100,000, an FPR of
0.001, and 100 Bloom filters, we reuse the primary key locations for all
Bloom filters to avoid repeated hash computations. As a result, the
blocked Bloom filter is also 5 times faster than the basic Bloom filter
in querying.
- Block BF TestLocation cost {"time": "529.97183ms"}
- Basic BF TestLocation cost {"time": "3.197430181s"}
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
See also #32642
`LocationCache` used map to store different locations for different K
which may cause lots of CPU time when get locations many times.
This PR change the implementation of LocationCache to store only the
location for the largest K used to totally remove the map access
operation.
See pprof from test of @XuanYang-cn
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #32165
There were some frequent scan in metacache:
- List all segments whose start positions not synced
- List compacted segments
Those scan shall cause lots of CPU time when flushed segment number is
large meanwhile `Flushed` segments can be skipped in those two scenarios
This PR make:
- Add segment state shortcut in metacache
- List start positions state before `Flushed`
- Make compacted segments state to be `Dropped` and use `Dropped` state
while scanning them
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #32642
This PR reuses hash locations for bloom filter prediction utilizing
`storage.Location`, like enhancement #32642.
Also adds a utility struct in storage: `LocationCache` to storage
locations for variable K (numbers of hash functions)
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
"-1" as `InvalidPartitionID` previously used as All partition place
holder in delete cases. It's confusing and hard to maintain when a const
var has more than one meaning.
This PR add `AllPartitionsID` to replace these usages in delete
scenarios.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #31143
This PR add short cut for datanoe metacache `WithSegmentIDs` filter,
which could just fetch segment from map with provided segmentIDs. Also
add benchmark for new implementation vs old one.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
This pr decoups importing segment from flush process by:
1. Exclude the importing segment from the flush policy, this approch
avoids notifying the datanode to flush the importing segment, which may
not exist.
2. When RootCoord call Flush, DataCoord directly set the importing
segment state to `Flushed`.
issue: https://github.com/milvus-io/milvus/issues/30359
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
See also #27675
BloomFilterSet.current shall be reset after RollStats, otherwise it will
keep tracking whole segment data causing the false positive ratio larger
than expected.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also: #27675
The bloom filter set initialized new BF with fixed configured `n`. This
value is always larger than the actual batch size and causes generated
BF using more memory.
This PR make write buffer to initialize BF with estimated batch size
from schema & configuration value.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #27675
Since serialization segment buffer does not related to sync manager can
shall be done before submit into sync manager. So that the pk statistic
file could be more accurate and reduce complex logic inside sync
manager.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #29092
`FlushSegments` transfer only `Growing` segment to flushing, if the
segment is in `Sealed` state before Datanode watch channel, the state
will never got satisfied for a segment be selected to be flushed.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #28736#28748
See also #27675
Previous PR: #28646
This PR fixes `SegmentNotFound` issue when compaction happens multiple
times and the buffer of first generation segment is sync due to stale
policy
Now the `CompactSegments` API of metacache shall update the compactTo
field of segmentInfo if the compactTo segment is also compacted to keep
the bloodline clean
Also, add the `CompactedSegment` SyncPolicy to sync the compacted
segment asap to keep metacache clean
Now the `SyncPolicy` is an interface instead of a function type so that
when it selects some segments to sync, we colud log the reason and
target segment
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #27675
Compacted segment info shall be removed after all buffer belongs to it
is sync-ed.
This PR add the cleanup function after triggerSyncTask logic:
- The buffer is stable and protected by mutex
- Cleanup fetches compacted & non-sync segment
- Remove segment info only there is no buffered maintained in manager
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #27675
- Fix LevelZero segment cannot be flushed
- Add level option for syncTask
- Invoke `AddSegment` when new LevelZero segment is allocated
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #27675
This PR make previously merged refactory of datanode go online
- Use write node to replace insert/delete node
- Use write buffer manager to control all buffers
- Use sync manager to control sync tasks instead of flush manager
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
