From: Antonin Houska <ah@cybertec.at>
To: pgsql-hackers@postgresql.org
Subject: Conflicting updates of command progress
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <117146.1744380740.1@localhost>
Date: Fri, 11 Apr 2025 16:12:20 +0200
Message-ID: <117147.1744380740@localhost>
Archived-At: <https://www.postgresql.org/message-id/117147.1744380740%40localhost>
Precedence: bulk

While working on [1] I realized that some field of pg_stat_progress_cluste has
weird value. On closer look I found out that cluster_rel() indirectly calls
index_build() and that overwrites the progress parameter. Following are the
parameter values in the master branch:

/* Phases of cluster (as advertised via PROGRESS_CLUSTER_PHASE) */
#define PROGRESS_CLUSTER_PHASE_SEQ_SCAN_HEAP	1
#define PROGRESS_CLUSTER_PHASE_INDEX_SCAN_HEAP	2
#define PROGRESS_CLUSTER_PHASE_SORT_TUPLES		3
#define PROGRESS_CLUSTER_PHASE_WRITE_NEW_HEAP	4
#define PROGRESS_CLUSTER_PHASE_SWAP_REL_FILES	5
#define PROGRESS_CLUSTER_PHASE_REBUILD_INDEX	6
#define PROGRESS_CLUSTER_PHASE_FINAL_CLEANUP	7

/* Progress parameters for CREATE INDEX */
/* 3, 4 and 5 reserved for "waitfor" metrics */
#define PROGRESS_CREATEIDX_COMMAND				0
#define PROGRESS_CREATEIDX_INDEX_OID			6
#define PROGRESS_CREATEIDX_ACCESS_METHOD_OID	8
#define PROGRESS_CREATEIDX_PHASE				9	/* AM-agnostic phase # */
#define PROGRESS_CREATEIDX_SUBPHASE				10	/* phase # filled by AM */
#define PROGRESS_CREATEIDX_TUPLES_TOTAL			11
#define PROGRESS_CREATEIDX_TUPLES_DONE			12
#define PROGRESS_CREATEIDX_PARTITIONS_TOTAL		13
#define PROGRESS_CREATEIDX_PARTITIONS_DONE		14
/* 15 and 16 reserved for "block number" metrics */

The only conflicting parameters I see here are
PROGRESS_CLUSTER_PHASE_REBUILD_INDEX vs PROGRESS_CREATEIDX_INDEX_OID (number
6). Fortunately, index_build() does not set PROGRESS_CREATEIDX_INDEX_OID so
there's no live bug. However [1] adds some PROGRESS_CLUSTER_ parameters, thus
making conflicts realistic.

AFAICS the current design does not consider that one progress-reporting
command can be called by another one. Not sure what the correct fix is. We can
either ignore update requests from the "nested" commands, or display the
progress of both. The latter approach is probably more invasive - is that
worth the effort?

[1] https://commitfest.postgresql.org/patch/5117/

-- 
Antonin Houska
Web: https://www.cybertec-postgresql.com