Here is a list of changes in numbers in the recent MySQL Cluster versions.
1) With the new changes of our checkpointing scheme the restart times in MySQL
Cluster are 3-4x faster for a database size of around 60 GByte.
2) The changes we have done in our parallel query execution parts have improved
some benchmark queries we have tested with up to a factor of 2.
3) A new wakeup threads in the NDB API means that each API node can deliver 10%
4) 7.6.6 introduces a new shared memory transporter to communicate between
colocated MySQL Server and NDB data nodes. This improves throughput by 10% and
can improve throughput on low thread counts as much as 60%.
5) Benchmarks using the Read Backup feature (released in 7.5) shows that
performance increases around 50% even in a network with a very high bandwidth.
6) A benchmark in the Oracle Cloud using the new feature to ensure locality of
reads within the same availability domain shows 50% improved throughput.
7) Parallelisation of the UNDO log applier speeds up this phase of the NDB recovery
by a factor of 5.
8) We have changed the defaults for index build during restarts that can speed up
the index build phase of the restart by a factor of 2.
9) We have changed the default batch sizes of index builds during restore of a
backup that can speed up restore by at least a factor of 2.
So as these numbers reflect MySQL Cluster 7.6 have significantly improved our
availability by decreasing the amount of time to restart a node. At the same
time we have also significantly improved performance of an SQL application
on top of MySQL Cluster using Read Backup feature, Fully Replicated feature,
a new shared memory transporter, a new feature for locality of reads in a
cloud environment and finally an improved throughput on each API node.
I will present those results in detail in the coming weeks. Some of those tests
was performed in the Oracle Cloud which means that anyone can reproduce the
exact same tests themselves which I think is a very nice feature with the