Is it really a good idea to restore to a previous copy of meta? Wouldn't it be better to fix meta using hbck2? The only thing we know for certain about a past meta state is it's in the past.
On Sat, Sep 12, 2020 at 9:27 AM Sean Busbey <bus...@apache.org> wrote: > Now that we have snapshots with TTLs, it would be helpful if we had > automated periodic snapshots of system tables that e.g. hbck2 could use to > do a forcible restore in cases of severe system damage. > > On Sat, Sep 12, 2020, 08:49 张铎(Duo Zhang) <palomino...@gmail.com> wrote: > > > When reviewing the snapshot related code, I found that we have some code > > for treating meta table. For example: > > > > > > > https://github.com/apache/hbase/blob/e5ca9adc54f9f580f85d21d38217afa97aa79d68/hbase-server/src/main/java/org/apache/hadoop/hbase/master/snapshot/TakeSnapshotHandler.java#L195 > > > > But when I tried to actually snapshot meta from client(in a UT), I got > the > > following error message: > > > > System table snapshots are not allowed > > > > It is thrown from this method: > > > > > > > https://github.com/apache/hbase/blob/e5ca9adc54f9f580f85d21d38217afa97aa79d68/hbase-client/src/main/java/org/apache/hadoop/hbase/snapshot/ClientSnapshotDescriptionUtils.java#L42 > > > > So I'm wondering whether we want to support snapshotting system tables or > > not. > > > > In general, at least for meta table, we can not disable it nor delete it, > > so even if we could support snapshotting it, we can not restore the > > snapshot... > > > > Thoughts? Thanks. > > >
