Details
-
Bug
-
Status: Open
-
Normal
-
Resolution: Unresolved
-
None
-
None
-
Normal
Description
I've run into some issues with data inconsistency in a situation where a single node is rejoining a 3-node cluster with RF=3. I'm running 3.7.
I have a client system which inserts data into a table with around 7 columns, named let's say A-F,id, and version. LWTs are used to make the inserts and updates.
Typically what happens is there's an insert of values id, V_a1, V_b1, ... , version=1, then another process will pick up rows with for example A=V_a1 and subsequently update A to V_a2 and version=2. Yet another process will watch for A=V_a2 to then make a second update to the same column, and set version to 3, with end result being <id, V_a3, V_b1, ... , V_f1, version=3> There's a secondary index on this A column (there's only a few possible values for A so not worried about the cardinality issue), though I've reproed with the new SASI index too.
If one of the nodes is down, there's still 2 alive for quorum so inserts can still happen. When I bring up the downed node, sometimes I get really weird state back which ultimately crashes the client system that's talking to Cassandra.
When reading I always select all the columns, but there is a conditional where clause that A=V_a2 (e.g. SELECT * FROM table WHERE A=V_a2). This read is for processing any rows with V_a2, and ultimately updating to V_a3 when complete. While periodically polling for A=V_a2 it is of course possible for the poller to to observe the old V_a2 value while the other parts of the client system process and make the update to V_a3, and that's generally ok because of the LWTs used for updates, an occassionaly wasted reprocessing run ins't a big deal, but when reading at serial I always expect to get the original values for columns that were never updated too. If a paxos update is in progress then I expect that completed before its value(s) returned. But instead, the read seems to be seeing the partial commit of the LWT, returning the old V_2a value for the changed column, but no values whatsoever for the other columns. From the example above, instead of getting <id, V_a3 V_b1, ... , version=3>, or even the older <id, V_a2, V_b1, ..., version=2> (either of which I expect and are ok), I get only <id, V_a2, version=2>, so the rest of the columns end up null, which I never expect. However this isn't persistent, Cassandra does end up consistent, which I see via sstabledump and cqlsh after the fact.
In my client system logs I record the insert / updates, and this inconsistency happens around the same time as the update from V_a2 to V_a3, hence my comment about Cassandra seeing a partial commit. So that leads me to suspect that perhaps due to the where clause in my read query for A=V_a2, perhaps one of the original good nodes already has the new V_a3 value, so it doesn't return this row for the select query, but the other good node and the one that was down still have the old value V_a2, so those 2 nodes return what they have. The one that was down doesn't yet have the original insert, just the update from V_a1 -> V_a2 (again I suspect, it's not been easy to verify), which would explain where <id, V_a2, version=2> comes from, that's all it knows about. However since it's a serial quorum read, I'd expect some sort of exception as neither of the remaining 2 nodes with A=V_a2 would be able to come to a quorum on the values for all the columns, as I'd expect the other good node to return <id, V_a2, V_b1, ..., version=2>
I know at some point nodetool repair should be run on this node, but I'm concerned about a window of time between when the node comes back up and repair starts/completes. It almost seems like if a node goes down the safest bet is to remove it from the cluster and rebuild, instead of simply restarting the node? However I haven't tested that to see if it runs into a similar situation.
It is of course possible to work around the inconsistency for now by detecting and ignoring it in the client system, but if there is indeed a bug I hope we can identify it and ultimately resolve it.
I'm also curious if this relates to CASSANDRA-12126, and also CASSANDRA-11219 may be relevant.
I've been reproducing with a combination of manually stopping one node, running a test script I have to trigger the client system to insert data, then manually restarting the node and waiting. It's consistently inconsistent, reproducing on most attempts
Summary timeline:
1. Shut down third node of three.
2. Insert <id, V_a1, V_b1, ... , version=1> (along with many others)
3. Start the third node. (start occurs concurrently with 4 & 5)
4. Update <id, V_a2, V_b1, ... , version=2> (along with others for which A=V_a1)
5. Update <id, V_a3, V_b1, ... , version=3> (along with many others for which A=V_a2)
5. Read
a. Expected: <id, V_a2, V_b1, ... , version=2> OR <id, V_a3, V_b1, ... , version=3>
b. Actual: <id, V_a2, version=2> // some fields are null