[ https://issues.apache.org/jira/browse/CASSANDRA-7066?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14523011#comment-14523011 ]
Benedict commented on CASSANDRA-7066: ------------------------------------- It looks like all of those could simply delegate to Directories.SSTableLister, which could concatenate all of the NEW transaction logs together to construct a filter. Directories.SSTableLister and CFS.scrubDataDirectories both work on state managed by c* itself, so we can expect the transaction logs to be maintained. So the only slight weird ones are CFS.loadNewSSTables and SSTableLoader.openSSTables, which would have interactions with users. Perhaps for user facing sstable construction we can keep TMP file suffixes? However in SSTableLoader, if the user isn't managing the directory carefully and is running with tmp files (or while sstable construction is happening directly to that directory) they're going to end up performing a lot of duplicated work, since we open every file in the directory. So they should expect _everything_ present to be sent, so I'm not convinced ignoring tmp files is the correct thing - we should probably be erroring out here and complaining that not every file can be opened. CFS.loadNewSSTables looks like it is for power users who have directly messed with the data directories, and I think we have to trust them to do it safely. They can use TMP file suffixes themselves as intermediary files if they want, and we'll just ignore them anyway since they will no longer match our filter. Tl;DR: I think we could keep both mechanisms for every action that isn't managed directly by C*, since there's no complexity cost, but I think I would rather see actual .tmp suffixes, and no special "Descriptor" for that, so that we just ignore the files. There's no atomicity requirement there, just safety for the end user. > Simplify (and unify) cleanup of compaction leftovers > ---------------------------------------------------- > > Key: CASSANDRA-7066 > URL: https://issues.apache.org/jira/browse/CASSANDRA-7066 > Project: Cassandra > Issue Type: Improvement > Components: Core > Reporter: Benedict > Assignee: Stefania > Priority: Minor > Labels: compaction > Fix For: 3.x > > > Currently we manage a list of in-progress compactions in a system table, > which we use to cleanup incomplete compactions when we're done. The problem > with this is that 1) it's a bit clunky (and leaves us in positions where we > can unnecessarily cleanup completed files, or conversely not cleanup files > that have been superceded); and 2) it's only used for a regular compaction - > no other compaction types are guarded in the same way, so can result in > duplication if we fail before deleting the replacements. > I'd like to see each sstable store in its metadata its direct ancestors, and > on startup we simply delete any sstables that occur in the union of all > ancestor sets. This way as soon as we finish writing we're capable of > cleaning up any leftovers, so we never get duplication. It's also much easier > to reason about. -- This message was sent by Atlassian JIRA (v6.3.4#6332)