pg_refclone.sh

#!/bin/bash
set -eux

ORIGINAL_DIR=$1
CLONE_DIR=$2

# create directory for clone
mkdir ${CLONE_DIR}

# make sure data directory cannot be started before we finish properly
cat <<EOT >> ${CLONE_DIR}/backup_label
pg_refclone.sh is not finished!
Removing this file and starting PG
will implicitly accept data corruption,
as the recovery will not run through
enough WAL to make things consistent.
EOT

# coprocess to handle backup control
coproc PSQL { psql -qtAX ; }
if [ $? -ne 0 ]; then exit 1; fi

# create replication slot so that we can later pull all the required WAL
echo "SELECT pg_create_physical_replication_slot('pg_refclone', TRUE);" >&${PSQL[1]}
if [ $? -ne 0 ]; then exit 1; fi
read -u ${PSQL[0]}

# drop slot later
clean_up () {
    echo "SELECT pg_drop_replication_slot('pg_refclone');" >&${PSQL[1]}
    echo '\q' >&${PSQL[1]}
}
trap clean_up EXIT

# start backup
echo "SELECT pg_backup_start('pg_refclone', TRUE);" >&${PSQL[1]}
if [ $? -ne 0 ]; then exit 1; fi
read -u ${PSQL[0]} START_LSN

echo $START_LSN

# do the reflink copy
cp -a --reflink=always ${ORIGINAL_DIR}/. ${CLONE_DIR}/

# make sure the clone is started in standby mode
touch ${CLONE_DIR}/standby.signal

# stop backup, saving return values for later use
# the return values are :lsn, :labelfile, :spcmapfile
echo "SELECT * FROM pg_backup_stop(FALSE) \gset" >&${PSQL[1]}
if [ $? -ne 0 ]; then exit 1; fi

echo "\echo :lsn"  >&${PSQL[1]}
read -u ${PSQL[0]} END_LSN
echo $END_LSN

# create backup_label file
echo "\o ${CLONE_DIR}/backup_label \qecho :labelfile" >&${PSQL[1]}

# create tablespace_map file
echo "\o ${CLONE_DIR}/tablespace_map \qecho :spcmapfile" >&${PSQL[1]}

# reset output
echo "\o" >&${PSQL[1]}

# get the WAL file name matching the backup end LSN, ensuring the latter is properly quoted
echo "SELECT pg_walfile_name(:'lsn');" >&${PSQL[1]}
read -u ${PSQL[0]} END_WALFILE
echo $END_WALFILE

# advance slot to start of backup (otherwise we get all the WAL produced before pg_backup_start was called)
echo "SELECT pg_replication_slot_advance('pg_refclone', '$START_LSN');" >&${PSQL[1]}
if [ $? -ne 0 ]; then exit 1; fi
read -u ${PSQL[0]}

# clean out pg_wal so pg_receivewal encounters no conflicts
rm -rf ${CLONE_DIR}/pg_wal
mkdir ${CLONE_DIR}/pg_wal
# pull all the WAL that we need, should end automatically when we reach backup end LSN
/usr/pgsql-18/bin/pg_receivewal -v -D ${CLONE_DIR}/pg_wal/ -S pg_refclone -E $END_LSN -s 1

# rename the WAL file containing the backup stop and SWITCH WAL record, as pg_receivewal would have kept them as ".partial" files.
mv ${CLONE_DIR}/pg_wal/${END_WALFILE}.partial ${CLONE_DIR}/pg_wal/${END_WALFILE}

# remove postmaster.pid file that was cloned, we are safe to run on our own now.
rm -f ${CLONE_DIR}/postmaster.pid