[HBASE-19260] Add lock back to avoid parallel accessing meta to locate region

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.3.1, 1.2.6, 2.0.0-alpha-3, 1.1.12
• Fix Version/s: 1.4.0, 1.3.2, 2.0.0-beta-1, 2.0.0, 1.2.7
• Hadoop Flags: Reviewed
• Assignee: Yu Li
• Reporter: Yu Li
• Created: 15/Nov/17 03:44
• Updated: 27/Jun/23 07:12
• Resolved: 17/Nov/17 22:31

Description

In branch-0.98 we have below codes to avoid accessing meta in parallel in HConnectionManager:

          Result regionInfoRow;
          // This block guards against two threads trying to load the meta
          // region at the same time. The first will load the meta region and
          // the second will use the value that the first one found.
          if (useCache) {
            if (TableName.META_TABLE_NAME.equals(parentTable) && usePrefetch &&
                getRegionCachePrefetch(tableName)) {
              synchronized (regionLockObject) {
                // Check the cache again for a hit in case some other thread made the
                // same query while we were waiting on the lock.
                ...
              }
            }
          ...

while in HBASE-10018 we removed such logic along with region-location-prefetching.

We regard this as an unexpected behavior change and observed below phenomenon in our product env:
1. Unnecessary connection setup to meta when multiple threads locating region in a client process
2. Priority handler of the RS holding meta region exhausted, application keep retrying and cause a vicious circle

To resolve this problem, we propose to add the userRegionLock back and keep the behavior in accordance with 0.98

Issue Links

• is broken by: HBASE-10018 Remove region location prefetching
• relates to: HBASE-27764 scan table is slow when enter hbase shell first time