fix ci port allocated (#17109)

### Summary
1. fix the issue of ci failures under specific circumstances.
2. fix ci  port allocated
3. modified unit_test file only run on python
This commit is contained in:
maoyifeng
2026-07-20 17:29:52 +08:00
committed by GitHub
parent 0e3a579fec
commit a150db6277

View File

@@ -190,7 +190,14 @@ jobs:
while IFS= read -r file; do
case "$file" in
# Docker, test, SDK, and workflow changes can alter either server path.
Dockerfile|docker/**|.github/workflows/**|test/**|sdk/**)
Dockerfile|docker/**|.github/workflows/**|sdk/**)
has_go=true
has_python=true
;;
test/unit_test/**)
has_python=true
;;
test/**)
has_go=true
has_python=true
;;
@@ -385,64 +392,124 @@ jobs:
# each container has its own /tmp and its own network namespace, and `ss` cannot
# see host ports bound by other runners' compose stacks. Use the shared daemon for
# both primitives:
# - offset mutex = a config-only docker network created atomically on the shared daemon
# without allocating a subnet from the daemon's address pools
# (docker network create is NOT idempotent: it fails on name conflict, which
# is exactly the cross-runner test-and-set we need);
# - per-port mutex = for each host port (base+offset) we intend to bind, create
# a docker network named ragflow-ci-port-<port> on the shared daemon. docker
# network create is NOT idempotent (fails on name conflict), so only one runner
# can claim a given port. This is per-PORT, not per-offset, because different
# (base, offset) pairs can produce the same host port (e.g. 6380+36497 ==
# 9380+33497 == 42877); a per-offset guard cannot prevent that collision.
# - port availability = docker ps published ports (sees every container on the
# shared daemon regardless of which runner started it).
# No file-system sharing required, no host netns required.
PORT_GUARD_PREFIX=ragflow-ci-portguard
port_offset_available() {
local offset=$1
local base port used
# Enumerate host-published tcp ports across ALL containers on the shared daemon.
# Format example: "0.0.0.0:38217->4222/tcp". We extract the left side.
used=$(sudo docker ps --format '{{.Ports}}' | grep -oE '[0-9]+->' | sed 's/->$//' | sort -u)
for base in "${PORT_BASES[@]}"; do
port=$((base + offset))
if grep -qx "${port}" <<< "${used}"; then
return 1
fi
done
return 0
}
cleanup_stale_port_locks() {
local now stale_after net ts
local now stale_after net ts before after
now=$(date -u +%s)
stale_after=$((2 * 60 * 60))
stale_after=$((60 * 60))
before=$(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}' | wc -l)
for net in $(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}'); do
ts=$(sudo docker network inspect --format '{{ index .Labels "ragflow-ci-created" }}' "${net}" 2>/dev/null || true)
if [[ "${ts}" =~ ^[0-9]+$ ]] && (( now - ts > stale_after )); then
# Remove if timestamp is non-numeric (legacy network from a previous code
# version) or if it is numeric and older than stale_after.
if [[ ! "${ts}" =~ ^[0-9]+$ ]] || (( now - ts > stale_after )); then
sudo docker network rm "${net}" >/dev/null 2>&1 || true
fi
done
after=$(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}' | wc -l)
echo " cleanup_stale_port_locks: ${before} -> ${after} guards (TTL=${stale_after}s)" >&2
}
reserve_port_offset() {
local attempt candidate guard
local attempt candidate base port guard g ok reject_reason create_err guard_subnet
local -a guards
local used guarded need_refresh used_n guarded_n debug_count create_fail_streak
cleanup_stale_port_locks
for attempt in $(seq 0 5999); do
echo "Reserving host ports for ${DOC_ENGINE} compose stack..." >&2
# Fail fast if the shared docker daemon is not reachable — avoids a
# long spin where every docker network create silently fails.
if ! sudo docker info >/dev/null 2>&1; then
echo " ERROR: 'sudo docker info' failed — docker daemon is not reachable from this runner" >&2
return 1
fi
echo " Running containers on shared daemon: $(sudo docker ps -q 2>/dev/null | wc -l)" >&2
echo " Docker networks on shared daemon: $(sudo docker network ls -q 2>/dev/null | wc -l)" >&2
need_refresh=1
debug_count=0
create_fail_streak=0
for attempt in $(seq 0 499); do
candidate=$(( PARTITION_BASE + ((GITHUB_RUN_ID + RUNNER_NUM * 1000 + attempt * 97) % PARTITION_SIZE) ))
guard="${PORT_GUARD_PREFIX}-${candidate}"
# Already held by some runner on this shared daemon.
if sudo docker network inspect "${guard}" >/dev/null 2>&1; then
# Refresh snapshots on first attempt and after a failed reservation.
if [[ ${need_refresh} -eq 1 ]]; then
used=$(sudo docker ps --format '{{.Ports}}' 2>/dev/null | grep -oE '[0-9]+->' | tr -d '>-' | sort -u || true)
# Strip both ragflow-ci-port- (current) and ragflow-ci-portguard- (legacy) prefixes.
guarded=$(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}' 2>/dev/null | sed 's/^ragflow-ci-port\(guard\)\?-//' | sort -u || true)
used_n=$(echo "${used}" | grep -c . 2>/dev/null || true)
guarded_n=$(echo "${guarded}" | grep -c . 2>/dev/null || true)
echo " Snapshot: ${used_n} bound ports, ${guarded_n} guarded ports" >&2
[[ ${used_n} -gt 0 ]] && echo " Bound sample: $(echo "${used}" | tr '\n' ' ' | cut -d' ' -f1-10)" >&2
[[ ${guarded_n} -gt 0 ]] && echo " Guarded sample: $(echo "${guarded}" | tr '\n' ' ' | cut -d' ' -f1-10)" >&2
need_refresh=0
fi
# Fast path: skip if any (base+candidate) port is already bound or guarded.
ok=1
reject_reason=""
for base in "${PORT_BASES[@]}"; do
port=$((base + candidate))
if grep -qx "${port}" <<< "${used}" 2>/dev/null; then
ok=0; reject_reason="port ${port} (base ${base}) already bound"; break
fi
if grep -qx "${port}" <<< "${guarded}" 2>/dev/null; then
ok=0; reject_reason="port ${port} (base ${base}) already guarded"; break
fi
done
if [[ ${ok} -eq 0 ]]; then
if [[ ${debug_count} -lt 5 ]]; then
echo " Candidate ${candidate}: rejected — ${reject_reason}" >&2
debug_count=$((debug_count + 1))
fi
continue
fi
if ! port_offset_available "${candidate}"; then
continue
fi
# Atomic cross-runner test-and-set: docker network create fails if the name
# already exists on the shared daemon, so even if two runners race here, only
# one wins; the loser retries the next candidate.
if sudo docker network create --config-only --label ragflow-ci-created="$(date -u +%s)" "${guard}" >/dev/null 2>&1; then
# Atomically claim every (base+candidate) host port via a per-port docker
# network. docker network create fails on name conflict, so only one runner
# can claim a given port. If any port is already claimed, undo partial
# reservations and try the next offset.
guards=()
ok=1
for base in "${PORT_BASES[@]}"; do
port=$((base + candidate))
guard="ragflow-ci-port-${port}"
# Use an explicit subnet from 10.0.0.0/8 (outside Docker's
# default 172.17-31.x.x pools) to bypass "all predefined
# address pools have been fully subnetted" errors on shared
# daemons where auto-allocation is exhausted.
guard_subnet="10.$((port / 256)).$((port % 256)).0/24"
if ! create_err=$(sudo docker network create --subnet "${guard_subnet}" --label ragflow-ci-created="$(date -u +%s)" "${guard}" 2>&1 >/dev/null); then
echo " docker network create ${guard} failed: ${create_err}" >&2
ok=0; break
fi
guards+=("${guard}")
done
if [[ ${ok} -eq 1 ]]; then
PORT_OFFSET=${candidate}
PORT_RESERVATION="${guard}"
PORT_RESERVATION="${guards[*]}"
echo "Reserved offset ${PORT_OFFSET} (attempt ${attempt})" >&2
return 0
fi
for g in "${guards[@]}"; do
sudo docker network rm "${g}" >/dev/null 2>&1 || true
done
# If docker network create keeps failing (daemon gone, permission, etc.),
# bail out early instead of looping 6000 times. A name-conflict failure
# should be rare because the fast path already filters guarded ports.
create_fail_streak=$((create_fail_streak + 1))
if [[ ${create_fail_streak} -ge 10 ]]; then
echo " docker network create failed ${create_fail_streak} consecutive candidates; aborting" >&2
return 1
fi
# Another runner just claimed a port; refresh snapshots before retrying.
need_refresh=1
done
echo " Exhausted ${attempt} candidates. Last snapshot: ${used_n} bound, ${guarded_n} guarded" >&2
return 1
}
@@ -745,9 +812,11 @@ jobs:
sudo docker rmi -f ${RAGFLOW_IMAGE}
fi
if [[ -n ${PORT_RESERVATION:-} ]]; then
# PORT_RESERVATION is now a docker network name on the shared daemon (created by the
# prepare step as a cross-runner port-offset mutex). Remove it to free the offset.
sudo docker network rm "${PORT_RESERVATION}" >/dev/null 2>&1 || true
# PORT_RESERVATION is a space-separated list of per-port docker network guards
# created by the prepare step. Remove each to free the ports for other runners.
for guard in ${PORT_RESERVATION}; do
sudo docker network rm "${guard}" >/dev/null 2>&1 || true
done
fi
@@ -896,64 +965,124 @@ jobs:
# each container has its own /tmp and its own network namespace, and `ss` cannot
# see host ports bound by other runners' compose stacks. Use the shared daemon for
# both primitives:
# - offset mutex = a config-only docker network created atomically on the shared daemon
# without allocating a subnet from the daemon's address pools
# (docker network create is NOT idempotent: it fails on name conflict, which
# is exactly the cross-runner test-and-set we need);
# - per-port mutex = for each host port (base+offset) we intend to bind, create
# a docker network named ragflow-ci-port-<port> on the shared daemon. docker
# network create is NOT idempotent (fails on name conflict), so only one runner
# can claim a given port. This is per-PORT, not per-offset, because different
# (base, offset) pairs can produce the same host port (e.g. 6380+36497 ==
# 9380+33497 == 42877); a per-offset guard cannot prevent that collision.
# - port availability = docker ps published ports (sees every container on the
# shared daemon regardless of which runner started it).
# No file-system sharing required, no host netns required.
PORT_GUARD_PREFIX=ragflow-ci-portguard
port_offset_available() {
local offset=$1
local base port used
# Enumerate host-published tcp ports across ALL containers on the shared daemon.
# Format example: "0.0.0.0:38217->4222/tcp". We extract the left side.
used=$(sudo docker ps --format '{{.Ports}}' | grep -oE '[0-9]+->' | sed 's/->$//' | sort -u)
for base in "${PORT_BASES[@]}"; do
port=$((base + offset))
if grep -qx "${port}" <<< "${used}"; then
return 1
fi
done
return 0
}
cleanup_stale_port_locks() {
local now stale_after net ts
local now stale_after net ts before after
now=$(date -u +%s)
stale_after=$((2 * 60 * 60))
stale_after=$((60 * 60))
before=$(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}' | wc -l)
for net in $(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}'); do
ts=$(sudo docker network inspect --format '{{ index .Labels "ragflow-ci-created" }}' "${net}" 2>/dev/null || true)
if [[ "${ts}" =~ ^[0-9]+$ ]] && (( now - ts > stale_after )); then
# Remove if timestamp is non-numeric (legacy network from a previous code
# version) or if it is numeric and older than stale_after.
if [[ ! "${ts}" =~ ^[0-9]+$ ]] || (( now - ts > stale_after )); then
sudo docker network rm "${net}" >/dev/null 2>&1 || true
fi
done
after=$(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}' | wc -l)
echo " cleanup_stale_port_locks: ${before} -> ${after} guards (TTL=${stale_after}s)" >&2
}
reserve_port_offset() {
local attempt candidate guard
local attempt candidate base port guard g ok reject_reason create_err guard_subnet
local -a guards
local used guarded need_refresh used_n guarded_n debug_count create_fail_streak
cleanup_stale_port_locks
for attempt in $(seq 0 5999); do
echo "Reserving host ports for ${DOC_ENGINE} compose stack..." >&2
# Fail fast if the shared docker daemon is not reachable — avoids a
# long spin where every docker network create silently fails.
if ! sudo docker info >/dev/null 2>&1; then
echo " ERROR: 'sudo docker info' failed — docker daemon is not reachable from this runner" >&2
return 1
fi
echo " Running containers on shared daemon: $(sudo docker ps -q 2>/dev/null | wc -l)" >&2
echo " Docker networks on shared daemon: $(sudo docker network ls -q 2>/dev/null | wc -l)" >&2
need_refresh=1
debug_count=0
create_fail_streak=0
for attempt in $(seq 0 499); do
candidate=$(( PARTITION_BASE + ((GITHUB_RUN_ID + RUNNER_NUM * 1000 + attempt * 97) % PARTITION_SIZE) ))
guard="${PORT_GUARD_PREFIX}-${candidate}"
# Already held by some runner on this shared daemon.
if sudo docker network inspect "${guard}" >/dev/null 2>&1; then
# Refresh snapshots on first attempt and after a failed reservation.
if [[ ${need_refresh} -eq 1 ]]; then
used=$(sudo docker ps --format '{{.Ports}}' 2>/dev/null | grep -oE '[0-9]+->' | tr -d '>-' | sort -u || true)
# Strip both ragflow-ci-port- (current) and ragflow-ci-portguard- (legacy) prefixes.
guarded=$(sudo docker network ls --filter "label=ragflow-ci-created" --format '{{.Name}}' 2>/dev/null | sed 's/^ragflow-ci-port\(guard\)\?-//' | sort -u || true)
used_n=$(echo "${used}" | grep -c . 2>/dev/null || true)
guarded_n=$(echo "${guarded}" | grep -c . 2>/dev/null || true)
echo " Snapshot: ${used_n} bound ports, ${guarded_n} guarded ports" >&2
[[ ${used_n} -gt 0 ]] && echo " Bound sample: $(echo "${used}" | tr '\n' ' ' | cut -d' ' -f1-10)" >&2
[[ ${guarded_n} -gt 0 ]] && echo " Guarded sample: $(echo "${guarded}" | tr '\n' ' ' | cut -d' ' -f1-10)" >&2
need_refresh=0
fi
# Fast path: skip if any (base+candidate) port is already bound or guarded.
ok=1
reject_reason=""
for base in "${PORT_BASES[@]}"; do
port=$((base + candidate))
if grep -qx "${port}" <<< "${used}" 2>/dev/null; then
ok=0; reject_reason="port ${port} (base ${base}) already bound"; break
fi
if grep -qx "${port}" <<< "${guarded}" 2>/dev/null; then
ok=0; reject_reason="port ${port} (base ${base}) already guarded"; break
fi
done
if [[ ${ok} -eq 0 ]]; then
if [[ ${debug_count} -lt 5 ]]; then
echo " Candidate ${candidate}: rejected — ${reject_reason}" >&2
debug_count=$((debug_count + 1))
fi
continue
fi
if ! port_offset_available "${candidate}"; then
continue
fi
# Atomic cross-runner test-and-set: docker network create fails if the name
# already exists on the shared daemon, so even if two runners race here, only
# one wins; the loser retries the next candidate.
if sudo docker network create --config-only --label ragflow-ci-created="$(date -u +%s)" "${guard}" >/dev/null 2>&1; then
# Atomically claim every (base+candidate) host port via a per-port docker
# network. docker network create fails on name conflict, so only one runner
# can claim a given port. If any port is already claimed, undo partial
# reservations and try the next offset.
guards=()
ok=1
for base in "${PORT_BASES[@]}"; do
port=$((base + candidate))
guard="ragflow-ci-port-${port}"
# Use an explicit subnet from 10.0.0.0/8 (outside Docker's
# default 172.17-31.x.x pools) to bypass "all predefined
# address pools have been fully subnetted" errors on shared
# daemons where auto-allocation is exhausted.
guard_subnet="10.$((port / 256)).$((port % 256)).0/24"
if ! create_err=$(sudo docker network create --subnet "${guard_subnet}" --label ragflow-ci-created="$(date -u +%s)" "${guard}" 2>&1 >/dev/null); then
echo " docker network create ${guard} failed: ${create_err}" >&2
ok=0; break
fi
guards+=("${guard}")
done
if [[ ${ok} -eq 1 ]]; then
PORT_OFFSET=${candidate}
PORT_RESERVATION="${guard}"
PORT_RESERVATION="${guards[*]}"
echo "Reserved offset ${PORT_OFFSET} (attempt ${attempt})" >&2
return 0
fi
for g in "${guards[@]}"; do
sudo docker network rm "${g}" >/dev/null 2>&1 || true
done
# If docker network create keeps failing (daemon gone, permission, etc.),
# bail out early instead of looping 6000 times. A name-conflict failure
# should be rare because the fast path already filters guarded ports.
create_fail_streak=$((create_fail_streak + 1))
if [[ ${create_fail_streak} -ge 10 ]]; then
echo " docker network create failed ${create_fail_streak} consecutive candidates; aborting" >&2
return 1
fi
# Another runner just claimed a port; refresh snapshots before retrying.
need_refresh=1
done
echo " Exhausted ${attempt} candidates. Last snapshot: ${used_n} bound, ${guarded_n} guarded" >&2
return 1
}
@@ -1251,7 +1380,9 @@ jobs:
sudo docker rmi -f ${RAGFLOW_IMAGE}
fi
if [[ -n ${PORT_RESERVATION:-} ]]; then
# PORT_RESERVATION is now a docker network name on the shared daemon (created by the
# prepare step as a cross-runner port-offset mutex). Remove it to free the offset.
sudo docker network rm "${PORT_RESERVATION}" >/dev/null 2>&1 || true
# PORT_RESERVATION is a space-separated list of per-port docker network guards
# created by the prepare step. Remove each to free the ports for other runners.
for guard in ${PORT_RESERVATION}; do
sudo docker network rm "${guard}" >/dev/null 2>&1 || true
done
fi