maria-developers team mailing list archive

[Kristian Nielsen <knielsen@xxxxxxxxxxxxxxx>]

Michael Widenius <monty@xxxxxxxxxxxx> writes:

> Why doing an insert + delete instead of one update?

This is described in more detail in my blog under "On crash-safe slave":

    http://kristiannielsen.livejournal.com/17008.html

The basic problem with just one UPDATE is that this causes row lock contention
when we get parallel replication, such as running group committed transactions
in parallel, or the parallel replication patch of dingqi/Taobao.

Suppose that we replicate transactions T1 and T2 in parallel. They both try to
UPDATE the replication state. One will go first, say T1, and lock the row,
causing T2 to have to wait. Only when T1 has committed will it release row
locks, allowing T2 to do the update and later commit.

So T1 and T2 can not commit in parallel. This makes group commit impossible,
which is a huge performance hit in some workloads.

Instead, the MDEV-26 design does a new insert for every change of the
replication state. This avoids row locks, and allows group commit to
work. Periodically we must then delete no longer needed rows, which can be
done in batches, or every commit, or whatever.

 - Kristian.