src/test/singlenode_isolation2/sql/checkpoint_dtx_info.sql

-- Scenario to test, CHECKPOINT getting distributed transaction
-- information between COMMIT processing time window
-- `XLogInsert(RM_XACT_ID, XLOG_XACT_DISTRIBUTED_COMMIT)` and
-- insertedDistributedCommitted(). `delayChkpt` protects this
-- case. There used to bug in placement of getDtxCheckPointInfo() in
-- checkpoint code causing, transaction to be committed on coordinator
-- and aborted on segments. Test case is meant to validate
-- getDtxCheckPointInfo() gets called after
-- GetVirtualXIDsDelayingChkpt().
--
-- Test controls the progress of COMMIT executed in session 1 and of
-- CHECKPOINT executed in the checkpointer process, with high-level
-- flow:
--
-- 1. session 1: COMMIT is blocked at start_insertedDistributedCommitted
-- 2. checkpointer: Start a CHECKPOINT and wait to reach before_wait_VirtualXIDsDelayingChkpt
-- 3. session 1: COMMIT is resumed
-- 4. checkpointer: CHECKPOINT is resumed and executes to keep_log_seg to finally introduce panic and perform crash recovery
--
-- Bug existed when getDtxCheckPointInfo() was invoked before
-- GetVirtualXIDsDelayingChkpt(), getDtxCheckPointInfo() will not
-- contain the distributed transaction in session1 whose state is
-- DTX_STATE_INSERTED_COMMITTED.  Therefore, after crash recovery, the
-- 2PC transaction that has been committed on coordinator will be
-- considered as orphaned prepared transaction hence is aborted at
-- segments. As a result the SELECT executed by session3 used to fail
-- because the twopcbug table only existed on the coordinator.
--
1: select gp_inject_fault_infinite('start_insertedDistributedCommitted', 'suspend', 1);
1: begin;
1: create table twopcbug(i int, j int);
1&: commit;
2: select gp_inject_fault_infinite('before_wait_VirtualXIDsDelayingChkpt', 'skip', 1);
33&: checkpoint;
2: select gp_inject_fault_infinite('keep_log_seg', 'panic', 1);
-- wait to make sure we don't resume commit processing before this
-- step in checkpoint
2: select gp_wait_until_triggered_fault('before_wait_VirtualXIDsDelayingChkpt', 1, 1);
-- reason for this inifinite wait is just to avoid test flake. Without
-- this joining step "1<" may see "COMMIT" sometimes or "server closed
-- the connection unexpectedly" otherwise. With this its always
-- "server closed the connection unexpectedly".
2: select gp_inject_fault_infinite('after_xlog_xact_distributed_commit', 'infinite_loop', 1);
2: select gp_inject_fault_infinite('start_insertedDistributedCommitted', 'resume', 1);
1<:
33<:
-- wait until coordinator is up for querying.
3: select 1;
3: select count(1) from twopcbug;