RE: Conflict detection and logging in logical replication

Dear Hou,

While playing with the 0003 patch (the patch may not be ready), I found that
when the insert_exists event occurred, both apply_error_count and insert_exists_count
was counted.

```
-- insert a tuple on the subscriber
subscriber =# INSERT INTO tab VALUES (1);

-- insert the same tuple on the publisher, which causes insert_exists conflict
publisher =# INSERT INTO tab VALUES (1);

-- after some time...
subscriber =# SELECT * FROM pg_stat_subscription_stats;
-[ RECORD 1 ]--------+------
subid                | 16389
subname              | sub
apply_error_count    | 16
sync_error_count     | 0
insert_exists_count  | 16
update_differ_count  | 0
update_exists_count  | 0
update_missing_count | 0
delete_differ_count  | 0
delete_missing_count | 0
stats_reset          |
```

Not tested, but I think this could also happen for the update_exists_count case,
or sync_error_count may be counted when the tablesync worker detects the conflict.

IIUC, the reason is that pgstat_report_subscription_error() is called in the
PG_CATCH part in start_apply() even after ReportApplyConflict(ERROR) is called.

What do you think of the current behavior? I wouldn't say I like that the same
phenomenon is counted as several events. E.g., in the case of vacuum, the entry
seemed to be separated based on the process by backends or autovacuum.
I feel the spec is unfamiliar in that only insert_exists and update_exists are
counted duplicated with the apply_error_count.

An easy fix is to introduce a global variable which is turned on when the conflict
is found.

Thought?

Best regards,
Hayato Kuroda
FUJITSU LIMITED