(doris) branch master updated: [fix](chore) path gc should consider tablet migration (#30095)

[yi...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

yiguolei pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/doris.git


The following commit(s) were added to refs/heads/master by this push:
     new 7feb5ea6ad7 [fix](chore) path gc should consider tablet migration (#30095)
7feb5ea6ad7 is described below

commit 7feb5ea6ad7a0623d85948eb86529a978f2d1a03
Author: zhannngchen <48...@users.noreply.github.com>
AuthorDate: Tue Jan 23 19:42:33 2024 +0800

    [fix](chore) path gc should consider tablet migration (#30095)
    
    Background:
    
    Migration will create new tablet in different DataDir, the old tablet will be moved to TabletManager::_shutdown_tablets.
    The migration task won't copy data in stale rowsets to new tablet, so after migration, the new tablet don't contains stale rowsets of old tablet
    The path GC process will check every path, to make sure if it's an useless tablet, or an useless rowset. If it is, will remove data of these tablets/rowsets
    The issue:
    
    When path GC got a stale rowset path from the data dir of old tablet, it extract the tablet id and rowset id
    Then it check if the tablet id exists in TabletManager, and the answer is YES!
    It got the tablet instance, which is the new tablet, then it check if the stale rowset id from the old tablet path exists in the new tablet instance, and got the answer NO.
    The path GC process treat the rowset as an useless rowset, since it can't find anyone holds reference to it, then delete the data of this stale rowset.
    But some query may still holds reference to this stale rowset, the deletion will cause query failure.
    Solution:
    
    The lifecycle of all rowsets in a shutdown tablet, should be related with the lifecycle of this tablet
    We need to differentiate the old tablet and the new one created by migration task, while performing path GC.
---
 be/src/olap/data_dir.cpp | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/be/src/olap/data_dir.cpp b/be/src/olap/data_dir.cpp
index e84ebd403b0..0f6d279e729 100644
--- a/be/src/olap/data_dir.cpp
+++ b/be/src/olap/data_dir.cpp
@@ -670,7 +670,14 @@ void DataDir::_perform_path_gc_by_tablet(std::vector<std::string>& tablet_paths)
             std::swap(*forward, *backward);
             continue;
         }
-        if (auto tablet = _engine.tablet_manager()->get_tablet(tablet_id); !tablet) {
+        auto tablet = _engine.tablet_manager()->get_tablet(tablet_id);
+        if (!tablet || tablet->data_dir() != this) {
+            if (tablet) {
+                LOG(INFO) << "The tablet in path " << path
+                          << " is not same with the running one: " << tablet->data_dir()->_path
+                          << "/" << tablet->tablet_path()
+                          << ", might be the old tablet after migration, try to move it to trash";
+            }
             _engine.tablet_manager()->try_delete_unused_tablet_path(this, tablet_id, schema_hash,
                                                                     path);
             --backward;
@@ -717,6 +724,12 @@ void DataDir::_perform_path_gc_by_rowset(const std::vector<std::string>& tablet_
             continue;
         }
 
+        if (tablet->data_dir() != this) {
+            // Current running tablet is not in same data_dir, maybe it's a tablet after migration,
+            // will be reclaimed in the next time `_perform_path_gc_by_tablet`
+            continue;
+        }
+
         bool exists;
         std::vector<io::FileInfo> files;
         auto st = io::global_local_filesystem()->list(path, true, &files, &exists);


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org