berth updates

This commit is contained in:
2026-09-16 19:22:16 -03:00
parent b3cf24e750
commit 907b616d1b
8 changed files with 399 additions and 17 deletions

View File

@@ -2,11 +2,23 @@
# The remote box — the other half of the inventory.
#
# Usage:
# ./host.sh status|ports|services
# ./host.sh status|ports|services announce a batch; run nothing
# ./host.sh snapshot [--admin] print the probe that would run
# ./host.sh snapshot --run [--admin] run it, save it, then check it
# ./host.sh check [<snapshot-dir>] is the box healthy? offline, never ssh
#
# Announces what it would run over `ssh $HOST` and does not run it. Always the
# ssh alias, never a hostname: there is no `Host mcrn.ar` block, so a bare
# hostname offers every agent key and trips MaxAuthTries.
# Always the ssh alias, never a hostname: there is no `Host mcrn.ar` block, so
# a bare hostname offers every agent key and trips MaxAuthTries.
#
# snapshot + check is rig's `make check` asked of a REMOTE machine. rig's check
# reports and instructs and never fixes; so does this. rig's check already reads
# /proc/meminfo through an override path so its branches can be exercised
# "against another machine's real numbers" — this is that idea taken the rest
# of the way: collect the other machine's numbers ONCE, over one announced ssh
# session, then answer every question from the saved copy. The analysis can
# then be re-run, tested against fixtures, and compared across days, and it
# never needs the box to be reachable — which, on a box that is failing, it
# may not be.
set -euo pipefail
cd "$(dirname "$0")"
@@ -31,6 +43,235 @@ announce_batch() {
echo "learned rather than scrolled past."
}
# Overridable so selftest can point check at an empty root. snapshot still
# refuses to write anywhere git does not confirm is ignored — a root outside
# the repo is refused, not honoured.
SNAP_ROOT="${SNAP_ROOT:-./.snapshots}"
snapshot() {
local run="" admin=""
for a in "$@"; do
case "$a" in
--run) run=1 ;;
--admin) admin=1 ;;
*) echo "snapshot: unknown argument '$a'" >&2; exit 1 ;;
esac
done
if [ -z "$run" ]; then
echo "would send, over 'ssh $HOST bash -s', exactly this file — ctrl/remote/probe.sh:"
echo
sed 's/^/ | /' ./remote/probe.sh
if [ -n "$admin" ]; then
echo
echo "and, over 'ssh $HOST_ADMIN bash -s', exactly this — ctrl/remote/probe-admin.sh:"
echo
sed 's/^/ | /' ./remote/probe-admin.sh
fi
echo
echo "read-only, and NOT run. What you read above is the file that is sent;"
echo "there is no second copy. To run it:"
echo " bash ctrl/host.sh snapshot --run${admin:+ --admin}"
return 0
fi
local dir; dir="$SNAP_ROOT/$ESTATE/$(date -u +%Y%m%dT%H%M%SZ)"
assert_ignored "$dir/raw" # before anything is written, not after
mkdir -p "$dir"
echo "ssh $HOST bash -s < ctrl/remote/probe.sh"
if ! ssh -o BatchMode=yes "$HOST" bash -s < ./remote/probe.sh > "$dir/raw" 2> "$dir/stderr"; then
echo " ! ssh $HOST exited non-zero — output so far is kept in $dir" >&2
sed 's/^/ /' "$dir/stderr" >&2
fi
split_sections "$dir/raw" "$dir"
if [ -n "$admin" ]; then
echo "ssh $HOST_ADMIN bash -s < ctrl/remote/probe-admin.sh"
ssh -o BatchMode=yes "$HOST_ADMIN" bash -s < ./remote/probe-admin.sh \
> "$dir/raw-admin" 2>> "$dir/stderr" || \
echo " ! ssh $HOST_ADMIN exited non-zero — see $dir/stderr" >&2
split_sections "$dir/raw-admin" "$dir"
fi
echo "saved: $dir"
echo
check "$dir"
}
# One file per `### name` section. An empty section still gets its file, so
# "the probe ran and found nothing" is distinguishable from "it never ran".
split_sections() {
awk -v d="$2" '/^### /{ f = d "/" $2; printf "" > f; next } f { print > f }' "$1"
}
# docker prints decimal units (kB, MB, GB). Approximate on purpose: these are
# for choosing between levers, not for accounting.
AWK_MB='function mb(s, v, u) {
v = s; u = s; sub(/[A-Za-z]+$/, "", v); sub(/^[0-9.]+/, "", u)
if (u == "B") return v / 1e6
if (u == "kB" || u == "KB") return v / 1e3
if (u == "MB") return v + 0
if (u == "GB") return v * 1e3
if (u == "TB") return v * 1e6
return 0
}'
check() {
local f="${1:-}"
if [ -z "$f" ]; then
# `|| true`: with no snapshots, ls fails, pipefail fails the pipeline and
# set -e exits silently — on exactly the path that must explain itself.
f=$(ls -1d "$SNAP_ROOT/$ESTATE"/*/ 2>/dev/null | sort | tail -1) || true
f=${f%/}
fi
# Exit 1 here, unlike every other report in berth: with no snapshot there is
# no report, and a check that prints nothing alarming because it looked at
# nothing is the vacuous pass selftest exists to catch.
if [ -z "$f" ] || [ ! -f "$f/meta" ]; then
echo "no snapshot for estate '$ESTATE'. Take one — read-only, printed before it runs:"
echo " make host snapshot"
exit 1
fi
local taken box age_min
taken=$(sed -n 1p "$f/meta"); box=$(sed -n 2p "$f/meta")
age_min=$(( ( $(date -u +%s) - $(date -u -d "$taken" +%s 2>/dev/null || date -u +%s) ) / 60 ))
echo "snapshot $f"
echo " $box, taken $taken"
if [ "$age_min" -ge 60 ]; then
echo " ! $((age_min / 60))h old — this is how the box WAS. Re-snapshot before acting on it."
fi
# ── disk ────────────────────────────────────────────────────────────────
echo
echo "disk"
local rows; rows=$(awk 'NR > 1 && NF >= 6 { p = $5; sub(/%/, "", p); printf "%s\t%d\t%d\t%d\n", $6, p, $4, $2 }' "$f/df")
local full=0 mnt pct avail total
while IFS=$'\t' read -r mnt pct avail total; do
[ -z "$mnt" ] && continue
local mark=" "; [ "$pct" -ge 85 ] && mark=" !"
printf "%s %-22s %3d%% %7d MB free of %7d MB\n" "$mark" "$mnt" "$pct" "$avail" "$total"
[ "$pct" -ge 95 ] && full=1
done <<< "$rows"
awk 'NR > 1 && NF >= 6 { p = $5; sub(/%/, "", p); if (p + 0 >= 85) printf " ! inodes %s at %s%% — space can be free and writes still fail\n", $6, $5 }' "$f/df_inodes"
# The filesystem docker writes to: the longest mount point that prefixes
# /var/lib/docker. Usually /, but a separate /var would otherwise be missed.
local dmnt dpct davail dtotal
IFS=$'\t' read -r dmnt dpct davail dtotal < <(
awk -F'\t' '{ m = $1; if (index("/var/lib/docker", m) == 1 && (m == "/" || substr("/var/lib/docker", length(m) + 1, 1) == "/") && length(m) > best) { best = length(m); row = $0 } } END { print row }' <<< "$rows")
if [ "$full" = 1 ]; then
echo
echo " A full disk fails WRITES, not reads — so it rarely presents as 'disk':"
echo " pages render logged-out but 500 logged-in (sessions, DB writes)"
echo " databases crash-loop on start (they must write to start)"
echo " pushes and uploads die partway through"
fi
# ── memory — rig's check.sh mb_of, recomputed here (✖ B8), pointed at the box's /proc/meminfo ──
echo
echo "memory"
mb_of() {
awk -v k="$1:" '$1 == k { printf "%d", $2 / 1024; found = 1 } END { if (!found) printf "0" }' "$f/meminfo"
}
local mtotal mavail swap_used
mtotal=$(mb_of MemTotal); mavail=$(mb_of MemAvailable)
swap_used=$(( $(mb_of SwapTotal) - $(mb_of SwapFree) ))
printf " available %d MB of %d MB\n" "$mavail" "$mtotal"
[ "$mtotal" -gt 0 ] && [ $(( mavail * 100 / mtotal )) -lt 10 ] && \
echo " ! under 10% available — expect the OOM killer before anything says 'memory'"
[ "$swap_used" -gt 0 ] && printf " ! %d MB in swap\n" "$swap_used"
# ── containers ──────────────────────────────────────────────────────────
echo
echo "containers"
awk -F'\t' 'NF >= 3 { n++; if ($2 == "running") r++; else printf " ! %-26s %-11s %s\n", $1, $2, $3 }
END { printf " %d of %d running\n", r, n }' "$f/containers"
awk -F'\t' 'NF >= 3 && $3 + 0 > 0 { n = $2; sub(/^\//, "", n); printf "%d\t%s\n", $3, n }' "$f/container_config" \
| sort -rn | head -5 | awk -F'\t' '{ printf " restarted %4d× %s\n", $1, $2 }'
local unbounded
unbounded=$(awk -F'\t' 'NF >= 6 && $5 == "json-file" && ($6 == "" || $6 == "<no value>") { n = $2; sub(/^\//, "", n); printf "%s ", n }' "$f/container_config")
if [ -n "$unbounded" ]; then
echo " ! json-file logs with NO max-size — each grows until the disk is full:"
echo "$unbounded" | fold -s -w 70 | sed 's/^/ /'
fi
# ── what would free space, cheapest first ───────────────────────────────
echo
echo "space — what would free it, and what it costs (≈MB; printed, never run)"
local dfrows; dfrows=$(awk "$AWK_MB"'
NR > 1 && NF >= 5 {
line = $0; gsub(/\([0-9]+%\)/, "", line); n = split(line, a, " ")
t = a[1]; for (i = 2; i <= n - 4; i++) t = t " " a[i]
printf "%s\t%d\t%d\n", t, mb(a[n - 1]), mb(a[n])
}' "$f/docker_df")
rec() { awk -F'\t' -v t="$1" '$1 == t { print $3 }' <<< "$dfrows"; }
local r_cache r_img r_ctr r_vol j_mb logs_mb=""
r_cache=$(rec "Build Cache"); r_img=$(rec "Images"); r_ctr=$(rec "Containers"); r_vol=$(rec "Local Volumes")
j_mb=$(awk '/take up/ { for (i = 1; i <= NF; i++) if ($i ~ /^[0-9.]+[KMGT]$/) { v = $i; u = substr(v, length(v)); sub(/[KMGT]$/, "", v); m = (u == "K" ? v / 1024 : u == "M" ? v : u == "G" ? v * 1024 : v * 1048576); printf "%d", m } }' "$f/journal" 2>/dev/null)
printf " %6s build cache docker builder prune free — nothing running uses it\n" "${r_cache:-?}"
printf " %6s unused images docker image prune -a next deploy re-pulls or rebuilds\n" "${r_img:-?}"
printf " %6s stopped ctrs docker container prune they cannot be 'start'ed again\n" "${r_ctr:-?}"
printf " %6s host journal sudo journalctl --vacuum-size=200M older host logs are gone\n" "${j_mb:-?}"
if [ -f "$f/admin_container_logs" ]; then
logs_mb=$(awk '$1 ~ /^[0-9]+$/ { s += $1 } END { print s + 0 }' "$f/admin_container_logs")
printf " %6s container logs sudo truncate -s 0 <log> RELIEF, not cure: they refill\n" "$logs_mb"
awk -F'\t' 'NR == FNR { id[substr($1, 1, 12)] = $2; next }
$1 ~ /^[0-9]+$/ && $1 >= 50 { n = $2; sub(/.*containers\//, "", n); k = substr(n, 1, 12); nm = (k in id ? id[k] : k); sub(/^\//, "", nm); printf " %6d %s\n", $1, nm }' \
"$f/container_config" "$f/admin_container_logs"
else
echo " ? container logs not sized: /var/lib/docker is root-only"
echo " — 'make host snapshot' with --admin sizes them"
fi
printf " %6s unused VOLUMES (no command) DATA. Unreferenced is not unwanted —\n" "${r_vol:-?}"
echo " a dangling volume can be the only copy. Decide per volume."
if [ -s "$f/registry_repos" ] && ! grep -q '^(not found' "$f/registry_repos"; then
local blobs; blobs=$(awk -F'\t' '$1 == "blobs_mb" { print $2 }' "$f/registry_repos")
printf " %6s registry blobs registry garbage-collect a DECISION — see registry below\n" "${blobs:-?}"
fi
# rig's check tells apart two failures with opposite fixes ("stopping that
# is what helps — a bigger VM would not"). The disk has the same pair.
if [ -n "$dmnt" ] && [ "${dtotal:-0}" -gt 0 ]; then
local deficit safe
deficit=$(( (dtotal - davail) - dtotal * 80 / 100 ))
safe=$(( ${r_cache:-0} + ${r_img:-0} + ${r_ctr:-0} + ${j_mb:-0} + ${logs_mb:-0} ))
echo
if [ "$deficit" -le 0 ]; then
printf " %s is under 80%% — no deficit.\n" "$dmnt"
elif [ "$safe" -ge "$deficit" ]; then
printf " %s needs ≈%d MB to reach 80%%; ≈%d MB is reclaimable without touching data.\n" "$dmnt" "$deficit" "$safe"
echo " Pruning is enough. A bigger disk would not be needed yet —"
echo " and if it refills, the cause is unbounded logs or builds, not size."
else
printf " %s needs ≈%d MB to reach 80%%; only ≈%d MB is reclaimable without touching data.\n" "$dmnt" "$deficit" "$safe"
echo " Pruning will NOT fix this. The rest is data — see the biggest directories."
echo " Grow the volume (EBS resizes online, then growpart + resize2fs), or move data off."
fi
fi
# ── where the space is ──────────────────────────────────────────────────
echo
echo "biggest directories (MB; unprivileged, so root-only trees are absent)"
awk -F'\t' 'NR <= 8 { printf " %7d %s\n", $1, $2 }' "$f/du_root" 2>/dev/null
echo " — under \$HOME:"
awk -F'\t' 'NR >= 2 && NR <= 8 { printf " %7d %s\n", $1, $2 }' "$f/du_home" 2>/dev/null
# ── registry ────────────────────────────────────────────────────────────
if [ -s "$f/registry_repos" ] && ! grep -q '^(not found' "$f/registry_repos"; then
echo
echo "registry (answers /v2/ anonymously — so read this list before collecting garbage)"
awk -F'\t' '$1 != "blobs_mb" { printf " %s %3s tag(s) %s\n", $1, $2, $3 }' "$f/registry_repos"
echo " A repository nobody recognises here is a finding, not clutter: with"
echo " anonymous push, it may be someone else's image in your deploy path."
fi
echo
echo "reports, never fixes. Every command above is printed; none was run."
}
guard_alias
case "${1:-status}" in
status)
@@ -59,5 +300,7 @@ for r in json.load(sys.stdin):
echo "stacks joining the shared 'gateway' network — intended design, but"
echo "nothing in the tree lists it. The inventory produces that list."
;;
*) echo "usage: $0 [status|ports|services]" >&2; exit 1 ;;
snapshot) shift; snapshot "$@" ;;
check) shift; check "$@" ;;
*) echo "usage: $0 [status|ports|services|snapshot [--run] [--admin]|check [<dir>]]" >&2; exit 1 ;;
esac