Details
-
Bug
-
Status: Resolved
-
Low
-
Resolution: Fixed
-
3.0.25, 3.11.11, 4.0.1, 4.1-alpha1, 4.1
-
None
Description
This is a follow up to/replacement of CASSANDRA-14073.
The behaviour that I want to avoid is someone trying to replace a node with the replace_address flag and mistakenly having that node listed in its own seed list which causes the node to generate a new set of random tokens before joining the ring.
Currently anytime an unbootstrapped node is listed in its own seed list and initial_token isn't set in the yaml, Cassandra will generate a new set of random tokens and join the ring regardless of whether it was replacing a previous node or not.
We could simply check for this configuration and refuse to start but I it's probably better (particularly for 3.0.X) if it's handled in the same manner as skipping streaming with the allow_unsafe_replace flag that was introduced in 3.X . This would still allow 3.0.X users the ability to re-bootstrap nodes without needing to re-stream all the data to the node again, which can be useful.
We currently handle replacing without streaming different;y between 3.0.X and 3.X. In 3.X we have the allow_unsafe_replace JVM flag to allow the use of auto_bootstrap: false in combination with the replace_address option. But in 3.0.X to perform the replacement of a node with the same IP address without streaming I believe you need to:
- Set replace_address (because the address is already in gossip)
- Include the node in its own seed list (to skip bootstrapping/streaming)
- Set the initial_token to the token/s owned by the previous node (to prevent it generating new tokens.
I believe if 3.0.X simply refused to start when a node has itself in its seed list and replace_address set this will completely block this operation.
Example patches to fix this edge case using allow_unsafe_replace:
Attachments
Issue Links
- supercedes
-
-
- Resolved
-