anybody hit SQLite "database is locked" errors?

anybody hit SQLite "database is locked" errors?

[Eric Wong]

Hey all, I've been reindexing frequently ahead of THREADED changes
while another process is doing grok-pull and triggering SQLite
reads in the post_update_hook.

This problem causes an exception in the --reindex process
because a read-only SQLite process is holding a SHARED lock.

In particular, the expensive (for LKML) `SELECT COUNT(*) FROM msgmap'
statement in our current examples/grok-pull.post_update_hook.sh
seems to be a culprit.  That's even opened with the sqlite3
`-readonly' flag, it still needs to acquire a SHARED lock in SQLite.

Other statements run by -httpd/-nntpd/-imapd may still cause
this, I think; but I haven't seen it...  Maybe others have?

WAL (write-ahead-log in SQLite) would be nice; but we can't use
it since we expect writers (-mda/-watch) to be a different Unix
user than -httpd/-nntpd/-imapd.

The Perl binding we use, DBD::SQLite sets it to 30000ms by
default, but the busy handler doesn't get fired in these cases.
See comments for sqlite3PagerSetBusyHandler in src/pager.c:
https://repo.or.cz/sqlite.git/blob_plain/33c269d09c0ec55e61cfe145b75e8cda59dc1d2d:/src/pager.c

(SQLite itself doesn't enable by busy handler by default, DBD::SQLite does)

I'm thinking of some possible solutions around it, but they're
fairly ugly and would make reads more expensive...  SQLite
itself seems saddled with retaining Windows 95 compatibility;
something we obviously won't have to care about.

[PATCH] grok-pull.post_update_hook: flock(2) before SQLite check

[Eric Wong]

Unlike DBD::SQLite, the sqlite3(1) CLI does not have a default
busy timeout enabled, so it easily times out while acquiring a
SHARED lock for read-only queries.  We can avoid battery-wasting
polling from the SQLite timeout handler by relying on flock(2)
as we do in our Perl code.

Furthermore, this avoids triggering some locking problems[1]
from a long "SELECT COUNT(*) ..." query and reindex.

While there may be other SQLite-related parallelism issues[1],
this works around one of them by relying on flock(2).

[1] https://public-inbox.org/meta/20200825001204.GA840@dcvr/
---
 examples/grok-pull.post_update_hook.sh | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/examples/grok-pull.post_update_hook.sh b/examples/grok-pull.post_update_hook.sh
index 1f51140f..77489472 100755
--- a/examples/grok-pull.post_update_hook.sh
+++ b/examples/grok-pull.post_update_hook.sh
@@ -25,11 +25,13 @@ then
 	inbox_dir=$(expr "$full_git_dir" : "$EPOCH2MAIN")
 	inbox_name=$(basename "$inbox_dir")
 	msgmap="$inbox_dir"/msgmap.sqlite3
+	inbox_lock="$inbox_dir"/inbox.lock
 else
 	inbox_fmt=1
 	inbox_dir="$full_git_dir"
 	inbox_name=$(basename "$inbox_dir" .git)
 	msgmap="$inbox_dir"/public-inbox/msgmap.sqlite3
+	inbox_lock="$inbox_dir"/ssoma.lock
 fi
 
 # run public-inbox-init iff unconfigured
@@ -118,7 +120,18 @@ esac
 # don't know what indexlevel a user wants
 if test -f "$msgmap"
 then
-	n=$(echo 'SELECT COUNT(*) FROM msgmap' | sqlite3 -readonly "$msgmap")
+	# We need to use flock(1) (from util-linux) to avoid timeouts
+	# and SQLite locking problems.
+	# FreeBSD has a similar lockf(1) utility, but it unlinks by
+	# default so we use `-k' to keep the lock on the FS.
+	FLOCK=flock
+	case $(uname -s) in
+	FreeBSD) FLOCK='lockf -k' ;;
+	# ... other OSes here
+	esac
+
+	n=$(echo 'SELECT COUNT(*) FROM msgmap' | \
+		$FLOCK $inbox_lock sqlite3 -readonly "$msgmap")
 	case $n in
 	0|'')
 		: v2 inboxes may be init-ed with an empty msgmap

Re: anybody hit SQLite "database is locked" errors?

[Eric Wong]

Eric Wong <e@80x24.org> wrote:
> In particular, the expensive (for LKML) `SELECT COUNT(*) FROM msgmap'
> statement in our current examples/grok-pull.post_update_hook.sh
> seems to be a culprit.  That's even opened with the sqlite3
> `-readonly' flag, it still needs to acquire a SHARED lock in SQLite.

Even without the proposed patch to the post_update_hook;
I haven't been able to reproduce the problem on a different
(faster) machine...