Yes, it's possible. A typical JVM GC pause for most configs is on the order of 50-200ms. If you have a host do a small collection/pause, then the read at #4 is basically racing the write at #1
(or, if you have an unhealthy cluster that's regularly dropping writes due to much larger problems, then it's even more likely) On Tue, Sep 1, 2020 at 12:10 AM Tobias Eriksson <tobias.eriks...@qvantel.com> wrote: > Hi > > We are seeing READ REPAIRs happening, and my understanding is this > > Setup 2 DCs with lots of Nodes, Replication Factor = 3 > > > > 1. Data written (on INSERT/UPDATE) > 2. Data replicated by Cassandra, but will not finish before (4) below > 3. Wait 80 ms on average > 4. Data read again with QUORUM i.e asking for atleast 2 out of 3 nodes > for result, and now ONE replies with inaccurate data > 5. (4) triggers a READ REPAIR > 6. The READ REPAIR now synchs to ALL nodes also in DC2 > > > > > > So my question is: Is it really possible that Cassandra within 80 ms is > not able to replicate to all 3 nodes in DC1 ? > > > > -Tobias >
