fix(duplication): fix plog do not gc after meta server restart #2066
Conversation
| @@ -701,6 +702,10 @@ void meta_duplication_service::recover_from_meta_state() | |||
| _meta_svc->get_meta_storage()->get_children( | |||
| get_duplication_path(*app), | |||
| [this, app](bool node_exists, const std::vector<std::string> &dup_id_list) { | |||
| dsn::defer([this, app]() { | |||
There was a problem hiding this comment.
The memory state will be updated [1] after get data from ZK successfully in line 721, why update it again?
There was a problem hiding this comment.
The memory state will be updated [1] after get data from ZK successfully in line 721, why update it again?
Because in function recover_from_meta_state, the vector dup_id_list may be empty.
Here is the detail of our online condition:
- Delete the existing hot standby first
A. Delete the corresponding dup from zk
B. Set duplicating in meta_server memory to 0 (not saved in zk), duplicating in zk is still 1 - meta_server restarts
Read dup and app_info data from zk to memory. At this time, dup is empty, but duplicating in zk is still 1 (because dup is empty, it is impossible to delete the hot standby to make duplicating 0). - Meta passes the hot standby status to the replica through
on_config_sync - Replica GC checks that the dup map is empty (
min_confirmed_decree = invalid_decree), but_is_duplication_master = true, so GC is not performed
There was a problem hiding this comment.
Thanks for the information. So this patch is aim to complete step 2? Why not complete the 1.B step?
There was a problem hiding this comment.
Thanks for the information. So this patch is aim to complete step 2? Why not complete the 1.B step?
Both two ways can fix this problem. In my opinion, the ways like step 2 is easier to implement.
There was a problem hiding this comment.
Correctness and consistent is important. Update the duplicating status to ZK is not very difficult if all duplications removed.
Or, if the duplicating memory status can be recovered from <app>/duplication ZK path, just competely remove the duplicating status from ZK.
What problem does this PR solve?
#2015
What is changed and how does it work?
When meta load duplication status from ZK, refresh
duplicatingstring in memory ,base on the value in ZK.Tests