Overview of the Issue

We use the table routing rules to decide where reads and writes are routed for tables which are part of MoveTables workflows.

For each table we have rules like

t1 => source.t1 
source.t1 => source.t1
t1@replica => source.t1@replica
source.t1@replica => source.t1@replica
target.t1 => source.t1

when no traffic is switched, or when all traffic is switched:

t1 => target.t1 
source.t1 => target.t1
t1@replica => target.t1@replica
source.t1@replica => target.t1@replica
target.t1 => target.t1

To determine whether writes are switched, we check the route for t1. If it is pointing to target it is considered that writes have been switched.

The problem comes the same table exists in different source keyspaces. For example if there are two workflows one from source1=>target1 and another from source2=>target2, both with a table t1.

Now if one workflow's writes are switched we end up with the route for t1 pointing to target1.t1. When the next workflow's writes are switched we now overwrite that rule so that it points to target2.t1. The logic now assumes writes for the first workflow have not been switched.

Reproduction Steps

.

Binary Version

c768b626575aa06025b347ea265d5cd4f3805fab

Operating System and Environment details

.

Log Fragments

No response