Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
Impala 2.8.0
-
None
Description
The following strange behavior is a result of the fix for IMPALA-3641. Please refer to IMPALA-3641 for an explanation of why this new bug here is preferred over IMPALA-3641.
The CREATE portion of a sequence of DROP IF EXISTS + CREATE from the same session may fail if there is a concurrent DROP [IF EXISTS] on the same object from a different session. Example and explanation:
- Thread A is executing a DROP TABLE t
- Thread B is executing a DROP TABLE IF EXISTS t
- Thread A drops the table first, gets the dropped object from the catalogd and applies the change to its local impalad (but not the impalad that thread B is connected to)
- Thread B executes DROP TABLE IF EXISTS t on the catalogd and finds that t is already gone, so an empty response is returned and no changes are applied to B's local impalad
- The impalad that B is connected to has not received a statestore update containing the DROP TABLE t executed by A
- B now executes CREATE TABLE t and finds that the table already exists
In general, a sequence of DROP+ADD from a single session is not guaranteed to always succeed if there are concurrent DROP/ADD operations of the same object from another session. However, this case is particularly strange because there is only a single ADD which was preceded by two DROPs, but the ADD still fails.