Corruption of table due to silent error in large table ALTER engine query from myisam

[Qubitium]

This bug maybe isolated to Percona mysql 5.7.latest stable. Did not test other mysql distro with rocksdb engine.

mysql error log:

2022-10-22T02:32:32.656099Z 5 [ERROR] Chunk size is too small to process merge.
2022-10-22T02:32:32.656137Z 5 [ERROR] Error during preparation of heap.

Large table migration from MYISAM to ROCKSDB via ALTER TABLE engine=rocksdb can lead to successful query but silently corrupted (empty) table that will never auto-recover.

After testing with various param/env changes, increasing rocksdb_merge_buf_size confirmed resolved this issue and bypasses this bug.

rocksdb_merge_buf_size = 2GB #64MB default
rocksdb_merge_combine_read_size = 8GB #1GB default
In summary, there are two critical bugs in this report that may cause silent table corruption via (ALTER engine=rocksdb).

1. BUG one: Rdb_index_merge::merge_heap_prepare() error is not correctly propagated to the ALTER sql query engine.

mysql-5.6/storage/rocksdb/rdb_index_merge.cc

Line 260 in 81ae98e

int Rdb_index_merge::merge_heap_prepare() {

2. BUG two: default rocksdb_merge_buf_size and rocksdb_merge_combine_read_size can cause fatal rocksdb migration/merging error.

3. Minor issue: L307 should also print the actual size that exceeded buffer to aid in debugging.

Cross posting bug from Percona jira which contains more details:

https://jira.percona.com/browse/PS-8440