topling / mytopling Goto Github PK

Show CompactionFilter stat in Rdb_compat_filter_factory WebView.

mtr test “rocksdb.add_index_inplace 'write_committed'” failed

output：
CURRENT_TEST: rocksdb.add_index_inplace
mysqltest: At line 411: Query 'ALTER TABLE t1 DROP INDEX idx, ADD UNIQUE KEY idx (j)' succeeded, should have failed with error '1062'

Feature: Show CompactionFilter stat in Rdb_tbl_prop_coll_factory WebView

show status like 'Rocksdb_git%'; shows none:

+------------------+------------------------------------------+
| Variable_name    | Value                                    |
+------------------+------------------------------------------+
| Rocksdb_git_date | none                                     |
| Rocksdb_git_hash | none                                     |
+------------------+------------------------------------------+

components: mytopling, toplingdb, sideplugin, cspp-memtab, cspp-wbwi, topling-rocks

in which topling-rocks is optional:

__attribute__((weak)) const char* git_version_hash_info_topling_rocks();
// ...
if (git_version_hash_info_topling_rocks) {
   // topling-rocks is present
}

MyRocks inplace_populate_sk() scan table one pass for each index, this waste CPUs, but this waste is not a hot spot because MyRocks MergeTree is very slow -- thus MergeTree is the hot spot.

MyTopling has auto_sort_sst, which is very fast, especially for building indexes, thus the wasted table scans are hot spot.

So creating multiple indexes shoule be realized by just one pass table scan, at least for non-unique index.

This needs decode SST smallest/largest userkey by mytopling data dictinary.

Now MyTopling inherited MyRocks's pack/unpack code, which is slow because it pack/unpack data dynamically, in table scan stage, pack/unpack consume , as show in this flame graph:

We can optimize it as below:

1. generate pack/unpack C++ code on create/alter table and create index
2. compile the code into shared lib
3. calling pack/unpack function in the shared lib

We expecting this will greatly improve pack/unpack performance.

InnoDB support parallel index creation by setting innodb_ddl_threads, mytopling should provide similar feature to speed up index creation and other long running ddl operation.

We had filed such a feature request to myrocks: facebook/mysql-5.6#1245
If myrocks community implemented such feature early, we can merge their contribution.