sankarh commented on a change in pull request #581: HIVE-21529 : Bootstrap ACID
tables as part of incremental dump.
URL: https://github.com/apache/hive/pull/581#discussion_r270926635
##########
File path: ql/src/java/org/apache/hadoop/hive/ql/exec/repl/ReplDumpTask.java
##########
@@ -193,27 +226,39 @@ private Long incrementalDump(Path dumpRoot, DumpMetaData
dmd, Path cmRoot, Hive
dmd.setDump(DumpType.INCREMENTAL, work.eventFrom, lastReplId, cmRoot);
dmd.write();
- // If external tables are enabled for replication and
- // - If bootstrap is enabled, then need to combine bootstrap dump of
external tables.
- // - If metadata-only dump is enabled, then shall skip dumping external
tables data locations to
- // _external_tables_info file. If not metadata-only, then dump the data
locations.
- if (conf.getBoolVar(HiveConf.ConfVars.REPL_INCLUDE_EXTERNAL_TABLES)
- && (!conf.getBoolVar(HiveConf.ConfVars.REPL_DUMP_METADATA_ONLY)
- || conf.getBoolVar(HiveConf.ConfVars.REPL_BOOTSTRAP_EXTERNAL_TABLES)))
{
+ // Examine all the tables if required.
+ if (shouldExamineTablesToDump()) {
Path dbRoot = getBootstrapDbRoot(dumpRoot, dbName, true);
+
+ // If we are bootstrapping ACID tables, stop all the concurrent
transactions and take a
+ // snapshot to dump those tables. Record the last event id in case we
are performing
+ // bootstrap of ACID tables.
+ String validTxnList = null;
+ long bootstrapLastReplId = 0;
+ if (conf.getBoolVar(HiveConf.ConfVars.REPL_BOOTSTRAP_ACID_TABLES)) {
+ validTxnList = getValidTxnListForReplDump(hiveDb);
+ bootstrapLastReplId =
hiveDb.getMSC().getCurrentNotificationEventId().getEventId();
Review comment:
There is a very corner case where this logic can be problem even for full
bootstrap.
Driver.java: L663
if ((queryState.getHiveOperation() != null)
&&
queryState.getHiveOperation().equals(HiveOperation.REPLDUMP)) {
setLastReplIdForDump(queryState.getConf());
}
openTransaction();
Here we capture last repl ID just before opening txn from REPL DUMP
execution thread (let's say T1)
If a concurrent thread (let's say T2) opens a txn after getting last repl ID
but before openTransaction() in T1. Let's say T2 writes in to the ACID table.
In this case, REPL DUMP would wait for T2 to commit txn before dumping ACID
table. Now, the snapshot of bootstrap dumped table includes the data written by
Thread T2. But, those events will be part of subsequent incremental dump.
When we apply these OpenTxn, allocateWriteId and commitTxn events, it may
create duplicate data.
Need some idempotent logic in replica side to handle this.
Please check if my theory make sense.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
