myrocks table lost all data after alter table in IngestExternalFile

The problem still exists for simple alter table x engine = rocksdb; queries.

Table data is lost entirely after a simple loop like:

The same is confirmed with a loop of alter table sbtest1 add column d int; alter table sbtest1 drop column d;

Big enough sysbench table (1M rows) and fast disk allow to reproduce in just 3-4 iterations.

Enabling rocksdb_bulk_load_use_sst_partitioner does not help this one.

Proposed a workaround, closing the issue.

Hi, we don’t know why rocksdb-bulk-load-use-sst-partitioner is disabled by default but I expect some side effects are possible because Meta is not afraid to change default values when it makes sense for them.

We received feedback from Meta. They suggested to use both:
1) a one line fix
2) set --rocksdb-bulk-load-use-sst-partitioner=1 option
In my experiments 1) doesn't solve the issue but 2) does.
Basically with 2) they gave us just a workaround. The problem is that rocksdb-bulk-load-use-sst-partitioner is disabled by default so I expect it has some drawbacks like e.g. worse performance.
Moreover Meta mentioned that they plan to rewrite bulk loading so probably the current implementation is buggy and hard to fix.

We noticed that rocksdb_stress.drop_cf_stress sometimes fails with the same issue i.e. Global seqno is required, but disabled.
I managed to simplify rocksdb_stress.drop_cf_stress and created rocksdb_stress.alter_table_crash. The crash is caused by assert(0) I put after Global seqno is required, but disabled is logged/encountered.

To trigger the issue each worker calls:

in the loop with different table names (tbl0X).

It is probably a race condition because I didn't manage to reproduce it with a single thread.

The issue is encountered without using rocksdb_bulk_load=1.

To reproduce the issue please use: 

I also confirmed that the issue started to appear after   was introduced to RocksDB repo.