[BUG] NPE happends when replay colocate join index

[wangbo]

code version: 0.13.11
Describe the bug

2021-02-18 10:27:29,961 ERROR (replayer|75) [EditLog.loadJournal():816] Operation Type 94
java.lang.NullPointerException: null
        at com.google.common.base.Preconditions.checkNotNull(Preconditions.java:878) ~[spark-dpp-1.0.0.jar:1.0.0]
        at org.apache.doris.catalog.ColocateTableIndex.replayAddTableToGroup(ColocateTableIndex.java:444) ~[palo-fe.jar:3.4.0]
        at org.apache.doris.persist.EditLog.loadJournal(EditLog.java:628) [palo-fe.jar:3.4.0]
        at org.apache.doris.catalog.Catalog.replayJournal(Catalog.java:2563) [palo-fe.jar:3.4.0]
        at org.apache.doris.catalog.Catalog$3.runOneCycle(Catalog.java:2347) [palo-fe.jar:3.4.0]
        at org.apache.doris.common.util.Daemon.run(Daemon.java:116) [palo-fe.jar:3.4.0]

From the above log we can see, when replay a ColocatePersistInfo, its table id can't be find in Database, then FE crash.

//master fe.log, the same table create successfully twice
// request 1:
2021-02-18 10:17:11,823 INFO (thrift-server-pool-983|56147) [Catalog.createOlapTable():3924] successfully create table[app_rt_rider_accept_waybill_rider_offset;58010577]
// request 2:
2021-02-18 10:17:25,256 INFO (thrift-server-pool-1200|155296) [Catalog.createOlapTable():3924] successfully create table[app_rt_rider_accept_waybill_rider_offset;58016408]

// master fe.log,  master receive create table forward request from the same follower
// request 1:
2021-02-18 10:16:45,427 DEBUG (thrift-server-pool-983|56147) [FrontendServiceImpl.forward():592] receive forwarded stmt 20766895 from FE: xxx
2021-02-18 10:16:45,427 INFO (thrift-server-pool-983|56147) [StmtExecutor.analyze():399] begin to analyze stmt: 5074185, forwarded stmt id: 20766895
// request 2:
2021-02-18 10:17:05,450 DEBUG (thrift-server-pool-1200|155296) [FrontendServiceImpl.forward():592] receive forwarded stmt 20766895 from FE: xxx
2021-02-18 10:17:05,451 INFO (thrift-server-pool-1200|155296) [StmtExecutor.analyze():399] begin to analyze stmt: 5074223, forwarded stmt id: 20766895

// follower fe.log, receive create table request from client
2021-02-18 10:16:45,426 INFO (doris-mysql-nio-pool-19139|472265) [StmtExecutor.analyze():399] begin to analyze stmt: 20766895, forwarded stmt id: 0
2021-02-18 10:16:45,427 INFO (doris-mysql-nio-pool-19139|472265) [MasterOpExecutor.forward():98] Forward statement 20766895 to Master TNetworkAddress(hostname:xxx, port:xxx)

From the log above, we can conclude the bug reason:
1 A Follower receive a create table request from client.
2 The Follower forward the request to Master.
3 Master response throw exception because of unknown reason, so the Follower try to forward twice.
4 Master request the same create table request twice, the second request create a new table, but didn't write table to Edit log. And ColocatePersistInfo write to Edit log twice.
5 When replay ColocatePersistInfo, it can't find its table id, then NPE happends;

[EmmyMiao87]

@wangbo

4 Master request the same create table request twice, the second request create a new table, but didn't write table to Edit log. And ColocatePersistInfo write to Edit log twice.

According to the fourth step you said, the edit log was not written when the table was built, and ColocatePersistInfo write to Edit log twice.
Doesn't this cause the table to be lost directly after fe restarts?

[wangbo]

@wangbo

4 Master request the same create table request twice, the second request create a new table, but didn't write table to Edit log. And ColocatePersistInfo write to Edit log twice.

According to the fourth step you said, the edit log was not written when the table was built, and ColocatePersistInfo write to Edit log twice.
Doesn't this cause the table to be lost directly after fe restarts?

    public boolean createTableWithLock(Table table, boolean isReplay, boolean setIfNotExist) {
        boolean result = true;
        writeLock();
        try {
            String tableName = table.getName();
            if (nameToTable.containsKey(tableName)) {
                result = setIfNotExist;
            } else {
                idToTable.put(table.getId(), table);
                nameToTable.put(table.getName(), table);

                if (!isReplay) {
                    // Write edit log
                    CreateTableInfo info = new CreateTableInfo(fullQualifiedName, table);
                    Catalog.getCurrentCatalog().getEditLog().logCreateTable(info);
                }
                if (table.getType() == TableType.ELASTICSEARCH) {
                    Catalog.getCurrentCatalog().getEsRepository().registerTable((EsTable)table);
                }
            }
            return result;
        } finally {
            writeUnlock();
        }
    }

The first request has already create the table successfully, which means the table info is written in edit log and memory.
But the second request won't write tablet info into memory and edit log if use specify if not exists in create stmt, because the table is already created successfully.
So the table won't be lost when restart.