Clik here to view.
Image may be NSFW.
Clik here to view.
When we run a SELECT in a replica server and it returns a different result to the one we would have obtained had we queried the source server instead, due to changes in the dataset that have not yet been replicated or synchronized to the replica, we get what is known as a stale read. My colleague Marcelo has blogged about Stale Reads and How to Fix Them before.
The need to avoid stale reads is a constraint that leads people into migrating to Percona XtraDB Cluster (PXC) / Galera Cluster or, more recently, Group Replication. I am focusing on PXC and Galera in this short blog post, which, similar to Group Replication, implements not exactly a synchronous replication model but what is commonly referred to as a “virtually synchronous” replication (but not without some polemic). In short, it means a transaction that is committed in one node is guaranteed (and there’s a certification process behind this) to have “reached” other nodes, but that doesn’t mean the changes this transaction carries with it have been applied on the other nodes already. Thus, stale reads can happen in PXC too. For the long version, see the manual page about Certification in Percona XtraDB Cluster.
Why does PXC implement virtually synchronous replication instead of fully synchronous replication? There are a few challenges to accomplish this but we can nail it down to performance constraints. Fully synchronous replication requires not only for the transactions to reach the other nodes but for the transactions to be applied to them too. And then they have to report back the success of the operation. For each transaction. This takes time and increases the overall commit time, which in practice makes writing to the database slow and impacts overall concurrency.
Evolution of “synchronous” replication on PXC
To be clear, no matter how you configure PXC, it won’t become a fully-synchronous system (neither will Group Replication, for that matter). BUT, it is possible to make it behave in such a way that it enforces data consistency. That’s what the variable wsrep_causal_reads has been used for. When this setting was introduced, it worked like a switch that one could use to kind of enable “synchronous” replication (as in data consistency) across the cluster – either for all transactions (globally) or at a connection basis (per session). In fact, what this setting does in practice is to enforce causality checks; quoting the manual: “When this variable is set to ON, the replica will wait until that event is applied before doing any other queries”. By waiting, it means blocking new writes until the expected consistency is reached.
Being able to enable “synchronous” replication for selected connections brought the potential to mix the best of both modes: operate with virtually synchronous replication by default (good performance) and use “synchronous” replication (once more, enforcing data consistency) when stale reads cannot be afforded.
The problem with this approach is that it requires some planning in advance to determine which connections should be configured with wsrep_causal_reads enabled. Without carefully planning ahead, what ends up happening in practice in many cases is having more connections configured with wsrep_causal_reads enabled than necessary. As a result, the cluster performance gets closer to that of operating with that setting applied globally.
The switch button that was wsrep_causal_reads evolved into the more granular wsrep_sync_wait, which allows the configuration of causality checks for different combinations of statements.
The cherry on top
While wsrep_sync_wait expanded the possibilities for causality checks, it still requires the “planning ahead” discussed above. I recently worked on the troubleshooting of a slow cluster and found it was configured with wsrep_sync_wait set to three (“Perform checks for READ, UPDATE, and DELETE statements”) all around, which is an even more “aggressive” configuration than the legacy wsrep_causal_reads. The justification for operating this way was that it was not possible to identify upfront which connection would be issuing a particular query that could not afford stale reads.
The good news is that starting with Percona XtraDB Cluster 8.0.26-16, which was released earlier this year, we can set wsrep_sync_wait on a per-query basis through the use of the optimizer hint SET_VAR: “The SET_VAR hint sets the session value of a system variable temporarily (for the duration of a single statement)”.
Here’s a simple example of employing SET_VAR to enforce the highest level of causality checks for a simple SELECT:
SELECT /*+ SET_VAR(wsrep_sync_wait=7) */ balance FROM accounts WHERE id=100;
If you find it is complicated to determine beforehand which connections should be configured to avoid stale reads, know that now you have the option to literally hard-code this constraint in your query. And if you end up deploying your database in a regular MySQL server, which is shipped without any wsrep_ settings, the optimizer hint is simply ignored and the query is processed normally.