Details
-
Bug
-
Status: Open
-
Minor
-
Resolution: Unresolved
-
3.1.3
-
None
-
None
-
CDH 6.3.2
Description
This is a Q&A about HMS. The questions are as follows:
Hive 3.0 introduces a metastore based on CachedStore. Is there a problem with phantom reading of RR-level records in the HMS cluster environment?
According to the description on the official website, our CachedStore is synchronized once every minute by default. In a cluster of multiple HMSs, if the table partition is modified concurrently, the cache will be inconsistent (I saw that ValidWriteIdList will also have short-term inconsistencies). Will there be an ABA problem for the rows in the table? ? It can be seen that the transaction in the code is at RR level, but the corresponding read operation does not include a `for update` statement.
For example, the function that modifies partitions, org.apache.hadoop.hive.metastore.RawStore#alterPartitions, will start a transaction, generate a directSql query partition (Mysql Locking Read is not used), and then modify mysql and mark the cache dirty. Suppose I have If transactions are written concurrently, will the metadata be messed up?
```
private List<Long> getPartitionIdsViaSqlFilter(
String catName, String dbName, String tblName, String sqlFilter,
List<? extends Object> paramsForFilter, List<String> joinsForFilter, Integer max)
throws MetaException {
...
String queryText =
"select " + PARTITIONS + ".\"PART_ID\" from " + PARTITIONS + ""
+ " inner join " + TBLS + " on " + PARTITIONS + ".\"TBL_ID\" = " + TBLS + ".\"TBL_ID\" "
+ " and " + TBLS + ".\"TBL_NAME\" = ? "
+ " inner join " + DBS + " on " + TBLS + ".\"DB_ID\" = " + DBS + ".\"DB_ID\" "
+ " and " + DBS + ".\"NAME\" = ? "
+ join(joinsForFilter, ' ')
+ " where " + DBS + ".\"CTLG_NAME\" = ? "
+ (StringUtils.isBlank(sqlFilter) ? "" : (" and " + sqlFilter)) + orderForFilter;
...
```