fix cron

2026-09-02 16:51:19 +07:00
parent f166f95b44
commit 99bc9665c3
6 changed files with 222 additions and 32 deletions
@@ -1,5 +1,5 @@
 DB_FILE=/data/db.sqlite
-CRONTAB="00 06 * * *"
+CRONTAB=00 06 * * *
 DELETE_AFTER=10
 KEEP_LAST=10
 TZ=Asia/Novosibirsk
@@ -19,8 +19,6 @@ RUN wget https://github.com/rclone/rclone/releases/download/v1.71.0/rclone-v1.71
    chown root:root /usr/bin/rclone && \
    chmod 755 /usr/bin/rclone

-RUN export TZ=/usr/share/zoneinfo/${TZ}
-
 RUN mkdir -p scripts
 COPY scripts/ /scripts

@@ -30,7 +30,7 @@ services:

 ```bash
 DB_FILE=/data/db.sqlite
-CRONTAB="00 06 * * *"
+CRONTAB=00 06 * * *
 DELETE_AFTER=10
 KEEP_LAST=10
 TZ=Asia/Novosibirsk
@@ -94,10 +94,38 @@ HEALTHCHECK_UUID=00000000-0000-0000-0000-000000000000

 Для self-hosted инстанса `HEALTHCHECK_URL` — это адрес его ping-эндпоинта, например `https://hc.domain.ru/ping`.

-## Как передаётся окружение в cron
+## Планировщик

-`crond` из busybox запускает джобы с пустым окружением, поэтому `entrypoint.sh`
-сохраняет env контейнера в `/tmp/container.env` (права 600), а cron-запись
-вызывает `/scripts/backup.sh --from-cron`, который его подгружает.
-Ручной запуск (`docker exec ... /scripts/backup.sh`) использует то окружение,
+`crond` из busybox не используется: под non-root uid (обычная ситуация в
+kubernetes с `securityContext.runAsUser`) он не загружает ни одного crontab и
+не сообщает об этом — в логе бесконечно идёт `wakeup dt=60`, а бекапов нет.
+Вместо него `entrypoint.sh` содержит свой планировщик на sh.
+
+Что это даёт:
+
+* работает под любым uid;
+* окружение наследуется напрямую, без выгрузки env в файл;
+* бекапы никогда не накладываются друг на друга — следующий запуск возможен
+  только после завершения предыдущего;
+* некорректный `CRONTAB` роняет контейнер на старте с внятным сообщением
+  (`[scheduler] FATAL: minute: '60' is outside 0-59`), а не приводит к тихому
+  простою.
+
+Синтаксис `CRONTAB` — стандартные пять полей: `*`, `5`, `1,15`, `2-6`, `*/15`,
+`8-18/2`, `30/5`. В месяцах и днях недели понимаются имена (`jan`, `mon`).
+Воскресенье — и `0`, и `7`. Если ограничены одновременно день месяца и день
+недели, запуск происходит при совпадении любого из них — как в `crontab(5)`.
+
+При старте контейнера бекап делается сразу, эта минута считается отработанной.
+
+Расписание видно в логе при запуске:
+
+```
+[scheduler] schedule: 30 07 * * * (TZ=Asia/Novosibirsk, now 2026-09-02 16:45:21 +0700)
+[scheduler] expanded: min=[30] hour=[7] dom=[1 2 3 ...] mon=[1 2 ...] dow=[0 1 ...] day-match=and
+```
+
+Ручной запуск: `docker exec ... /scripts/backup.sh` — использует то окружение,
 с которым его вызвали.
+
+Период опроса часов — 20 секунд, меняется через `POLL_INTERVAL` (секунды).
@@ -1 +1 @@
-1.2.2
+1.3.0
@@ -1,12 +1,5 @@
 #!/bin/sh

-# crond gives the job a bare environment (HOME/PATH/SHELL only), so the cron
-# entry passes --from-cron and the env dumped by entrypoint.sh is restored here.
-# A manual run keeps the environment it was invoked with.
-if [ "${1:-}" = "--from-cron" ] && [ -f /tmp/container.env ]; then
-  . /tmp/container.env
-fi
-
 set -u

 BACKUP_FILE="sqlite_$(date "+%F-%H%M%S").tar.gz"
@@ -1,22 +1,193 @@
 #!/bin/sh

-set -e
+# Planner for periodic backups.
+#
+# busybox crond is deliberately not used here: when the container runs under a
+# non-root uid (a very common k8s securityContext) crond loads no crontab at
+# all and never says why - it just keeps printing "wakeup dt=60" forever while
+# no backup is ever made. This loop does the same job, works under any uid and
+# fails loudly on a bad schedule.

-# crond runs jobs with a bare environment (HOME/PATH/SHELL only), so the
-# container env is dumped here and sourced back by backup.sh on every run.
-ENV_FILE=/tmp/container.env
-export -p | grep -v -E "^export (PWD|SHLVL|OLDPWD|_)=" > "$ENV_FILE"
-chmod 600 "$ENV_FILE"
+# -f matters: the schedule is full of '*' and word splitting would otherwise
+# expand them against the files in the working directory.
+set -euf

-# Create crontab in a writable location and set proper permissions
-mkdir -p /tmp/crontabs
-echo "${CRONTAB:-"0 * * * *"} /scripts/backup.sh --from-cron >> /proc/1/fd/1 2>&1" > /tmp/crontabs/root
-chmod 644 /tmp/crontabs/root
+SCHEDULE_RAW="${CRONTAB:-0 * * * *}"
+POLL_INTERVAL="${POLL_INTERVAL:-20}"

-echo "[entrypoint] schedule: ${CRONTAB:-"0 * * * *"} (TZ=${TZ:-UTC})"
+log()  { echo "[scheduler] $*"; }
+die()  { echo "[scheduler] FATAL: $*" >&2; exit 1; }

-# Run initial backup (do not abort the container if it fails)
-/scripts/backup.sh || echo "[entrypoint] initial backup failed, continuing to crond"
+# A value coming from .env, a configmap or a helm chart may carry quotes or a
+# trailing CR; both silently break the schedule, so strip them up front.
+SCHEDULE=$(printf '%s' "$SCHEDULE_RAW" | tr -d '\r' \
+  | sed -e 's/^"\(.*\)"$/\1/' -e "s/^'\(.*\)'\$/\1/")

-# Start crond in foreground with debug output
-exec crond -f -c /tmp/crontabs -d 0
+MONTH_NAMES="jan feb mar apr may jun jul aug sep oct nov dec"
+DOW_NAMES="sun mon tue wed thu fri sat"
+
+# Resolve a three-letter month/weekday name to its number, echo the input back
+# when it is not a name.
+resolve_name() { # $1=token $2=names $3=base
+  _tok=$(echo "$1" | tr '[:upper:]' '[:lower:]')
+  _n=0
+  for _w in $2; do
+    if [ "$_tok" = "$_w" ]; then
+      echo $((_n + $3))
+      return 0
+    fi
+    _n=$((_n + 1))
+  done
+  echo "$1"
+}
+
+# Drop a leading zero so that "07" is not read as an invalid octal number.
+strip_zero() {
+  _v=${1#0}
+  [ -n "$_v" ] || _v=0
+  echo "$_v"
+}
+
+# Expand one crontab field into the explicit list of values it matches.
+# Supports *, N, a-b, a,b, */n, a-b/n and N/n.
+expand_field() { # $1=spec $2=lo $3=hi $4=field name $5=names $6=names base
+  _spec=$1; _lo=$2; _hi=$3; _name=$4; _names=$5; _base=$6
+  _out=''
+  _oifs=$IFS
+  IFS=,
+  for _part in $_spec; do
+    IFS=$_oifs
+
+    _step=1
+    case $_part in
+      */*)
+        _step=${_part##*/}
+        _part=${_part%%/*}
+        ;;
+    esac
+    case $_step in
+      ''|*[!0-9]*) die "$_name: bad step in '$_spec'" ;;
+    esac
+    _step=$(strip_zero "$_step")
+    [ "$_step" -ge 1 ] || die "$_name: step must be >= 1 in '$_spec'"
+
+    case $_part in
+      '*')
+        _s=$_lo; _e=$_hi
+        ;;
+      *-*)
+        _s=$(resolve_name "${_part%%-*}" "$_names" "$_base")
+        _e=$(resolve_name "${_part##*-}" "$_names" "$_base")
+        ;;
+      *)
+        _s=$(resolve_name "$_part" "$_names" "$_base")
+        # "N/step" means "from N to the end of the range", like vixie cron.
+        if [ "$_step" -gt 1 ]; then _e=$_hi; else _e=$_s; fi
+        ;;
+    esac
+
+    for _v in "$_s" "$_e"; do
+      case $_v in
+        ''|*[!0-9]*) die "$_name: '$_part' is not a number or a known name" ;;
+      esac
+    done
+    _s=$(strip_zero "$_s")
+    _e=$(strip_zero "$_e")
+    [ "$_s" -ge "$_lo" ] && [ "$_e" -le "$_hi" ] \
+      || die "$_name: '$_part' is outside ${_lo}-${_hi}"
+    [ "$_s" -le "$_e" ] || die "$_name: reversed range '$_part'"
+
+    _i=$_s
+    while [ "$_i" -le "$_e" ]; do
+      _out="$_out $_i"
+      _i=$((_i + _step))
+    done
+
+    IFS=,
+  done
+  IFS=$_oifs
+  [ -n "$_out" ] || die "$_name: empty field"
+  echo "$_out"
+}
+
+in_list() { # $1=list $2=value
+  case " $1 " in
+    *" $2 "*) return 0 ;;
+  esac
+  return 1
+}
+
+# Word splitting is the point here, and globbing is off (set -f).
+# shellcheck disable=SC2086
+set -- $SCHEDULE
+[ $# -eq 5 ] || die "CRONTAB must have 5 fields, got $#: '$SCHEDULE_RAW'"
+SPEC_MIN=$1 SPEC_HOUR=$2 SPEC_DOM=$3 SPEC_MON=$4 SPEC_DOW=$5
+
+LIST_MIN=$(expand_field  "$SPEC_MIN"  0 59 minute  '' 0)
+LIST_HOUR=$(expand_field "$SPEC_HOUR" 0 23 hour    '' 0)
+LIST_DOM=$(expand_field  "$SPEC_DOM"  1 31 day     '' 0)
+LIST_MON=$(expand_field  "$SPEC_MON"  1 12 month   "$MONTH_NAMES" 1)
+LIST_DOW=$(expand_field  "$SPEC_DOW"  0  7 weekday "$DOW_NAMES"   0)
+# Sunday is both 0 and 7.
+if in_list "$LIST_DOW" 7 && ! in_list "$LIST_DOW" 0; then
+  LIST_DOW="$LIST_DOW 0"
+fi
+
+# crontab(5): when both day-of-month and day-of-week are restricted, a run
+# happens if either of them matches, not both.
+if [ "$SPEC_DOM" != '*' ] && [ "$SPEC_DOW" != '*' ]; then
+  DAY_MATCH=or
+else
+  DAY_MATCH=and
+fi
+
+log "schedule: $SCHEDULE (TZ=${TZ:-UTC}, now $(date '+%F %T %z'))"
+log "expanded: min=[${LIST_MIN# }] hour=[${LIST_HOUR# }] dom=[${LIST_DOM# }] mon=[${LIST_MON# }] dow=[${LIST_DOW# }] day-match=$DAY_MATCH"
+
+matches_now() { # $1=minute $2=hour $3=dom $4=month $5=dow
+  in_list "$LIST_MIN"  "$1" || return 1
+  in_list "$LIST_HOUR" "$2" || return 1
+  in_list "$LIST_MON"  "$4" || return 1
+
+  if [ "$DAY_MATCH" = or ]; then
+    in_list "$LIST_DOM" "$3" || in_list "$LIST_DOW" "$5" || return 1
+  else
+    in_list "$LIST_DOM" "$3" || return 1
+    in_list "$LIST_DOW" "$5" || return 1
+  fi
+  return 0
+}
+
+TERMINATE=0
+trap 'TERMINATE=1' TERM INT
+
+# Initial backup: do not abort the container if it fails, the schedule matters
+# more than this one run.
+/scripts/backup.sh || log "initial backup failed, continuing to the schedule"
+
+# The startup backup counts as this minute's run, otherwise a schedule that
+# matches the minute the container came up would immediately back up twice.
+LAST_RUN=$(date '+%Y-%m-%dT%H:%M')
+while [ "$TERMINATE" -eq 0 ]; do
+  # Polling instead of sleeping to the minute boundary: a backup can take
+  # minutes and the clock can jump, LAST_RUN is what keeps a minute from
+  # firing twice.
+  # shellcheck disable=SC2046
+  set -- $(date '+%Y-%m-%dT%H:%M %M %H %d %m %w')
+  KEY=$1
+  MIN=$(strip_zero "$2"); HOUR=$(strip_zero "$3")
+  DOM=$(strip_zero "$4"); MON=$(strip_zero "$5"); DOW=$(strip_zero "$6")
+
+  if [ "$KEY" != "$LAST_RUN" ] && matches_now "$MIN" "$HOUR" "$DOM" "$MON" "$DOW"; then
+    LAST_RUN=$KEY
+    log "triggered at $KEY"
+    /scripts/backup.sh || log "backup failed, waiting for the next run"
+  fi
+
+  # Backgrounded so that SIGTERM is handled right away instead of after the
+  # whole interval.
+  sleep "$POLL_INTERVAL" &
+  wait $! || true
+done
+
+log "terminating"
@@ -1 +1 @@
 .2.2
 .3.0