-
Bug
-
Resolution: Won't Do
-
Minor
-
3.6.4
-
None
-
Workaround Exists
-
I'm not 100% sure, but it seems like there might be a defect here.
I'm using TCP, S3_PING, and MERGE3.
I've set logical_addr_cache_max_size to 2 for testing purposes, although I don't think the value of this setting affects my test results.
I start a single node, node A. Then I start a second node, node B.
I then repeatedly shutdown and restart node B.
Each time node B starts, a new row is added to the .list file stored in S3.
But even if I continue this process for 15 minutes, old rows are never removed from the .list file, so it continues to grow in size.
I've read the docs and mailing list threads, so I'm aware that the list is not immediately updated as soon as a member leaves. But I was expecting that when a view change occurs, nodes no longer in the view would be marked for removal (line 2193 of TP.java) and then after the logical_addr_cache_expiration has been reached and the reaper kicks in, once a new node joins, the expired cache entries would be purged from the file.
I dug in to the code a bit, and what seems to be happening is that the MERGE3 protocol periodically generates a FIND_MBRS event. S3_PING retrieves the membership from the .list file, which includes expired nodes. And then all of these members are re-added to the logical address cache (line 157 of S3_PING.java, line 533 of Discovery.java, line 2263 of TP.java).
So expired nodes are continually re-added to the logical address cache, preventing them from ever being reaped.