MB-36424 to 6.0.4
When couchstore writes out modified node elements during document saves, it it supposed to limit the number of bytes written in a single node to a _chunk_threshold - by default 1279 bytes. If the node is larger than the limit it should be split into multiple sibling nodes.
However, this limit is not respected, resulting in overly-large nodes being written out. In the case of the by-seqno B-Tree (which always writes values to the rightmost leaf as seqnos are increasing), it results in all leaf elements residing in a single leaf node. Moreover, this means that adding another element to the B-Tree effectively re-writes the entire tree, resulting in massive Write Amplification.
Steps to Reproduce
- Start 2-node cluster run (single replica):
- Start a SyncWrite workload; single threaded client updating the same 100,000 items (each key 10 times) with level=persistMajority:
- Observe the op/s and Write Amplification
- Op/s should be broadly constant (given both key and seqno B-Tree should have a constant number of items in them).
- Write amplification should also be broadly constant.
- The op/s quickly drops from a peak of ~600 down to 150:
- The Write Amplification increase (corresponding with the top in op/s) from 5.6x, up to 743x .
- Both op/s and Write Amplification temporarily recover when compaction occurs, but same pattern is observed over time: