Wednesday, January 31, 2018

MySQL Cluster 7.6.4 is out

MySQL Cluster 7.6.4 DMR is out.

This new version contains a number of goodies.

1) Local checkpoint algorithm have been rewritten
The new checkpointing is designed to scale to at least 16 TBytes of DataMemory sizes
Checkpoints will be much faster, this decreases recovery times significantly
Table fragments that are not updated will not need any new checkpoints written
Checkpoint size on disk is significantly decreased

2) MySQL Cluster Configurator (MCC, Auto Installer)
MCC is significantly improved. Particularly for setups where you
have external computers either on-premise or in the cloud.

3) New cloud feature
In the cloud with availability domains/zones it is possible to have
10x difference between latency inside an AD compared to between
ADs. To run MySQL Cluster in a cloud region with synchronous
replication between ADs one can now configure nodes with a
LocationDomainId. This LocationDomainId will be used to
ensure that transaction coordinator is placed in the same AD and
that we always prefer reading data from our own AD if possible.

4) New ODirectSyncFlag
When using ODirect there are a number of file systems that ensures that writes
are also synched to disk. If the user knows that he is working in such an
environment setting ODirectSyncFlag can improve disk write speeds by
around 2-3x. This is particularly interesting when using hard drives.

5) Change default behaviour of restart configuration
We changed the BuildIndexThreads from 0 to 128 to improve speed of
index rebuilds. We added a new configuration setting to specify which
CPUs that can be used for index rebuilds.

We increased batch sizes (and made them configurable) to improve
performance of unique index creation and online add node and some
other algorithms.

We changed the default algorithm for initial node restart to rebuild indexes
in a special phase.

All these changes can lead to a very significant reduction in restore times.

6) Many improvements to our parallel query implementation (pushdown join,
SPJ). The improvement depends on the queries, but in our special benchmark
query we have improved latency of query execution to almost half.

7) Parallel UNDO log applier for disk columns
The phase where we apply the UNDO log is now fully parallelised over all
LDM threads. For a scenario with 4 LDM threads we've seen a speed up of
5x for the UNDO log applier phase (this is only used to UNDO changes in
pages containing disk columns during a restart).

8) Bug fixes
We have continued our quality improvements to ensure that each new version
is even more stable compared to the previous one.

No comments: