docs: prove sessions are shared by PostgreSQL, not Infinispan replication

Experiment 0 with three probes: cross-node refresh/logout, cache counter deltas around a single login, and cache entry ownership. Each node caches only what it handled; cache totals sum exactly to the database count.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
DongHyeonka
2026-09-04 09:58:10 +09:00
co-authored by Claude Opus 5
parent 006da7d490
commit e5ebaeb623
20 changed files with 1114 additions and 0 deletions
+55
View File
@@ -0,0 +1,55 @@
#!/usr/bin/env bash
# Experiment 0c — where does a session entry actually live?
#
# Experiment 0b showed keycloak-1's session cache never moved when keycloak-0
# handled a login. That leaves two explanations:
#
# (a) a DISTRIBUTED cache with owners=1 — entries are spread across nodes by
# consistent hashing, and this one happened to land on keycloak-0;
# (b) a LOCAL cache — each node only ever caches what it handled itself.
#
# They are distinguished by driving logins at the OTHER node. Under (a) the
# entries would keep landing on both nodes regardless of who was asked. Under
# (b) the count rises only on the node that received the request.
set -uo pipefail
NS="${NS:-keycloak-lab}"
N="${N:-5}"
K0_IP=$(kubectl -n "$NS" get pod keycloak-0 -o jsonpath='{.status.podIP}')
K1_IP=$(kubectl -n "$NS" get pod keycloak-1 -o jsonpath='{.status.podIP}')
ADMIN_PW=$(kubectl -n "$NS" get secret keycloak-lab-secrets \
-o jsonpath='{.data.KC_BOOTSTRAP_ADMIN_PASSWORD}' | base64 -d)
echo "수집 시각: $(date '+%Y-%m-%d %H:%M:%S %Z')"
echo " keycloak-0 = $K0_IP ($(kubectl -n "$NS" get pod keycloak-0 -o jsonpath='{.spec.nodeName}'))"
echo " keycloak-1 = $K1_IP ($(kubectl -n "$NS" get pod keycloak-1 -o jsonpath='{.spec.nodeName}'))"
echo
kubectl -n "$NS" run kc-own --rm -i --restart=Never \
--image=curlimages/curl:8.11.1 --quiet --command -- sh -c "
O=/tmp/o; : > \$O
ent() {
curl -s --retry 3 --max-time 20 http://\$1:9000/metrics \
| grep -E '^vendor_statistics_approximate_entries_unique.cache=.sessions' \
| awk '{print \$NF}'
}
login() { i=0; while [ \$i -lt $N ]; do
curl -s -o /dev/null -X POST http://\$1:8080/realms/master/protocol/openid-connect/token \
-d grant_type=password -d client_id=admin-cli \
-d username=admin -d 'password=$ADMIN_PW'
i=\$((i+1)); done; sleep 5; }
{
printf '%-32s %12s %12s\n' '단계' 'k0 entries' 'k1 entries'
printf '%-32s %12s %12s\n' '시작' \"\$(ent $K0_IP)\" \"\$(ent $K1_IP)\"
login $K1_IP
printf '%-32s %12s %12s\n' 'keycloak-1 에 로그인 ${N}회' \"\$(ent $K0_IP)\" \"\$(ent $K1_IP)\"
login $K0_IP
printf '%-32s %12s %12s\n' 'keycloak-0 에 로그인 ${N}회' \"\$(ent $K0_IP)\" \"\$(ent $K1_IP)\"
} >> \$O
cat \$O
" 2>&1 | grep -v '^pod .* deleted$'
echo
echo "=== 대조: PostgreSQL 에는 몇 건인가 ==="
kubectl -n "$NS" exec deploy/postgres -- psql -U keycloak -d keycloak -tAc \
"select count(*) from offline_user_session where offline_flag='0'" 2>/dev/null | sed 's/^/ online 세션 /'
+83
View File
@@ -0,0 +1,83 @@
#!/usr/bin/env bash
# Experiment 0b — does the Infinispan cache itself replicate, or do both nodes
# merely agree because they read the same database?
#
# Experiment 0 proved the two nodes give the same answers. That alone does NOT
# prove Infinispan replicated anything: with persistent-user-sessions (the
# Keycloak 26 default) the session is written to PostgreSQL, so two nodes reading
# one database would agree even with the cache disabled entirely.
#
# This script separates the two by measuring the cache counters on BOTH nodes
# around a single login. If the write on keycloak-0 shows up as cache activity
# on keycloak-1, the replication is real and not a database artifact.
set -uo pipefail
NS="${NS:-keycloak-lab}"
K0_IP=$(kubectl -n "$NS" get pod keycloak-0 -o jsonpath='{.status.podIP}')
K1_IP=$(kubectl -n "$NS" get pod keycloak-1 -o jsonpath='{.status.podIP}')
ADMIN_PW=$(kubectl -n "$NS" get secret keycloak-lab-secrets \
-o jsonpath='{.data.KC_BOOTSTRAP_ADMIN_PASSWORD}' | base64 -d)
echo "수집 시각: $(date '+%Y-%m-%d %H:%M:%S %Z')"
echo
# 파드 출력을 스트리밍으로 받으면 조각이 유실된다. 실제로 첫 시도에서
# keycloak-1 의 스냅샷과 그 다음 마커가 통째로 사라져 델타가 0 으로 보였다.
# 파드 안에서 파일로 모았다가 마지막에 한 번만 내보낸다.
kubectl -n "$NS" run kc-delta --rm -i --restart=Never \
--image=curlimages/curl:8.11.1 --quiet --command -- sh -c "
set -u
K0='http://$K0_IP'; K1='http://$K1_IP'
O=/tmp/o.txt; : > \$O
snap() {
curl -s --retry 3 --retry-connrefused --max-time 20 \$1:9000/metrics \
| grep -E '^vendor_(statistics_(stores|hits|misses|approximate_entries_unique)|rpc_manager_replication_count)\{cache=\"(sessions|clientSessions)\"' \
| sed 's/,cache_manager=\"keycloak\"//; s/,node=\"[^\"]*\"//' >> \$O
}
echo '###BEFORE_K0' >> \$O; snap \$K0
echo '###BEFORE_K1' >> \$O; snap \$K1
echo '###LOGIN' >> \$O
curl -s -o /dev/null -w 'http_code=%{http_code}\n' -X POST \
\"\$K0:8080/realms/master/protocol/openid-connect/token\" \
-d grant_type=password -d client_id=admin-cli \
-d username=admin -d 'password=$ADMIN_PW' >> \$O
sleep 5
echo '###AFTER_K0' >> \$O; snap \$K0
echo '###AFTER_K1' >> \$O; snap \$K1
echo '###END' >> \$O
cat \$O
" 2>&1 | grep -v '^pod .* deleted$' > /tmp/cache-delta.txt
python3 - /tmp/cache-delta.txt <<'PY'
import re, sys
raw = open(sys.argv[1]).read()
blocks, cur = {}, None
for line in raw.splitlines():
if line.startswith('###'):
cur = line[3:]; blocks[cur] = {}
elif cur and '{' in line:
m = re.match(r'(\S+?)\{cache="(\w+)"\}\s+(\S+)', line)
if m:
blocks[cur][(m.group(1), m.group(2))] = float(m.group(3))
print('=== 로그인은 keycloak-0 에만 보냈다 ===')
code = [l for l in raw.splitlines() if l.startswith('http_code=')]
print(' 로그인 응답: ' + (code[0] if code else '없음'))
for n in ('BEFORE_K0','BEFORE_K1','AFTER_K0','AFTER_K1'):
if not blocks.get(n):
print(f' !! {n} 스냅샷이 비었다 — 델타를 신뢰할 수 없다')
print()
hdr = f" {'계수기':<42} {'캐시':<15} {'전':>8} {'후':>8} {'증가':>7}"
for node in ('K0', 'K1'):
who = 'keycloak-0 (로그인을 받은 노드)' if node == 'K0' else 'keycloak-1 (아무 요청도 받지 않은 노드)'
print(f'=== {who} ===')
print(hdr)
b, a = blocks.get(f'BEFORE_{node}', {}), blocks.get(f'AFTER_{node}', {})
for k in sorted(set(b) | set(a)):
before, after = b.get(k[0:2], 0.0), a.get(k[0:2], 0.0)
d = after - before
mark = ' ←' if d else ''
name = k[0].replace('vendor_statistics_', '').replace('vendor_rpc_manager_', 'rpc.')
print(f" {name:<42} {k[1]:<15} {before:>8.0f} {after:>8.0f} {d:>+7.0f}{mark}")
print()
PY
+207
View File
@@ -0,0 +1,207 @@
#!/usr/bin/env bash
# Experiment 0 — is a session created on one Keycloak node usable on the other?
#
# Forming a cluster is not the same as sharing session state. The Infinispan log
# says "cluster view (2)", but that only proves the members found each other.
#
# Design notes, learned the hard way:
#
# * Every probe has a CONTROL. A result from the far node means nothing unless
# the same call against the issuing node is also measured. The first version
# of this script reported "403 on keycloak-1" as if it were a replication
# failure; the issuing node returned 403 too, and the cause was a missing
# openid scope. Measure both, always.
#
# * Sessions are tracked by SID, not by count. Both the test login and the
# admin API calls create sessions for the same user, so counts are noisy.
# A specific session id either appears in a node's answer or it does not.
#
# * The probe is the REFRESH TOKEN grant, not userinfo. userinfo only validates
# a signature and can succeed on a node that knows nothing about the session.
# Refreshing requires the node to find the session, check it is alive, and
# write back a new refresh time — it actually touches the session store.
#
# Talks to pod IPs directly: going through nginx/Traefik would hide which node
# handled each request, which is the entire question.
#
# ./deploy/lab/scripts/experiment-session-replication.sh
set -uo pipefail
NS="${NS:-keycloak-lab}"
OUT="${OUT:-/tmp/session-replication}"
mkdir -p "$OUT"
PSQL="kubectl -n $NS exec deploy/postgres -- psql -U keycloak -d keycloak -tAc"
echo "수집 시각: $(date '+%Y-%m-%d %H:%M:%S %Z')"
echo
K0_IP=$(kubectl -n "$NS" get pod keycloak-0 -o jsonpath='{.status.podIP}')
K1_IP=$(kubectl -n "$NS" get pod keycloak-1 -o jsonpath='{.status.podIP}')
K0_NODE=$(kubectl -n "$NS" get pod keycloak-0 -o jsonpath='{.spec.nodeName}')
K1_NODE=$(kubectl -n "$NS" get pod keycloak-1 -o jsonpath='{.spec.nodeName}')
ADMIN_PW=$(kubectl -n "$NS" get secret keycloak-lab-secrets \
-o jsonpath='{.data.KC_BOOTSTRAP_ADMIN_PASSWORD}' | base64 -d)
echo "=== 대상 ==="
printf ' keycloak-0 %-14s %s\n' "$K0_IP" "$K0_NODE"
printf ' keycloak-1 %-14s %s\n' "$K1_IP" "$K1_NODE"
echo
echo "=== [0] 실험 전 DB 세션 ==="
$PSQL "select offline_flag, count(*) from offline_user_session group by offline_flag" 2>/dev/null \
| sed 's/^/ offline_flag=/' || echo " (없음)"
echo
# 파드 하나 안에서 전 단계를 실행한다. 단계마다 파드를 새로 띄우면 토큰을
# 단계 사이로 넘길 수 없다.
kubectl -n "$NS" run kc-probe --rm -i --restart=Never \
--image=curlimages/curl:8.11.1 --quiet --command -- sh -c "
set -u
K0='http://$K0_IP:8080'; K1='http://$K1_IP:8080'
TOKEN_EP='/realms/master/protocol/openid-connect/token'
jget() { sed -n \"s/.*\\\"\$1\\\":\\\"\\([^\\\"]*\\)\\\".*/\\1/p\"; }
# ── [1] keycloak-0 에서 로그인. 이 노드가 세션의 출생지다 ──────────────────
LOGIN=\$(curl -s -X POST \"\$K0\$TOKEN_EP\" \
-d grant_type=password -d client_id=admin-cli \
-d username=admin -d 'password=$ADMIN_PW')
echo '###STEP1_LOGIN'; echo \"\$LOGIN\"
AT=\$(echo \"\$LOGIN\" | jget access_token)
RT=\$(echo \"\$LOGIN\" | jget refresh_token)
# ── [2] 관리 API 조회용 토큰. 세션 오염을 피하려고 따로 하나만 더 만든다 ──
ADMTOK=\$(curl -s -X POST \"\$K0\$TOKEN_EP\" \
-d grant_type=password -d client_id=admin-cli \
-d username=admin -d 'password=$ADMIN_PW' | jget access_token)
CID=\$(curl -s -H \"Authorization: Bearer \$ADMTOK\" \
\"\$K0/admin/realms/master/clients?clientId=admin-cli\" | jget id | head -1)
# ── [3] 두 노드에 같은 질문을 한다: admin-cli 의 세션 목록 ────────────────
echo '###STEP3_SESSIONS_K0'
curl -s -H \"Authorization: Bearer \$ADMTOK\" \
\"\$K0/admin/realms/master/clients/\$CID/user-sessions?max=100\"
echo
echo '###STEP3_SESSIONS_K1'
curl -s -H \"Authorization: Bearer \$ADMTOK\" \
\"\$K1/admin/realms/master/clients/\$CID/user-sessions?max=100\"
echo
# ── [4] 대조군: keycloak-0 이 발급한 refresh token 을 keycloak-0 에 쓴다 ──
# 먼저 반대편에 써야 하므로 여기서는 쓰지 않고, 순서를 [5] 뒤로 미룬다.
# refresh token 은 회전(rotation)되므로 한 번 쓰면 옛 것이 무효가 된다.
# 따라서 '반대편 먼저'가 유일하게 의미 있는 순서다.
# ── [5] 시험군: keycloak-0 이 발급한 refresh token 을 keycloak-1 에 쓴다 ──
echo '###STEP5_REFRESH_ON_K1'
curl -s -w '\nhttp_code=%{http_code}\n' -X POST \"\$K1\$TOKEN_EP\" \
-d grant_type=refresh_token -d client_id=admin-cli -d \"refresh_token=\$RT\"
RT2=\$(curl -s -X POST \"\$K1\$TOKEN_EP\" \
-d grant_type=refresh_token -d client_id=admin-cli -d \"refresh_token=\$RT\" \
| jget refresh_token)
# ── [6] 무효화가 반대 방향으로도 전파되는가 ───────────────────────────────
# keycloak-1 에서 로그아웃시키고, keycloak-0 에서 갱신을 시도한다.
echo '###STEP6_LOGOUT_VIA_K1'
curl -s -o /dev/null -w 'http_code=%{http_code}\n' -X POST \"\$K1/realms/master/protocol/openid-connect/logout\" \
-d client_id=admin-cli -d \"refresh_token=\$RT2\"
echo '###STEP7_REFRESH_ON_K0_AFTER_LOGOUT'
curl -s -w '\nhttp_code=%{http_code}\n' -X POST \"\$K0\$TOKEN_EP\" \
-d grant_type=refresh_token -d client_id=admin-cli -d \"refresh_token=\$RT2\"
echo '###END'
" > "$OUT/raw.txt" 2>&1
sed -i '/^pod .* deleted$/d' "$OUT/raw.txt"
python3 - "$OUT/raw.txt" <<'PY' | tee "$OUT/report.txt"
import base64, json, sys
raw = open(sys.argv[1]).read()
blocks, cur = {}, None
for line in raw.splitlines():
if line.startswith('###'):
cur = line[3:]; blocks[cur] = []
elif cur is not None:
blocks[cur].append(line)
get = lambda k: '\n'.join(blocks.get(k, [])).strip()
def j(s):
try: return json.JSONDecoder().raw_decode(s.strip())[0]
except Exception: return None
def claims(tok):
p = tok.split('.')[1]; p += '=' * (-len(p) % 4)
return json.loads(base64.urlsafe_b64decode(p))
login = j(get('STEP1_LOGIN'))
if not login or 'access_token' not in login:
print('로그인 실패:', get('STEP1_LOGIN')[:300]); sys.exit(1)
ac = claims(login['access_token'])
rc = claims(login['refresh_token'])
SID = ac['sid']
print('=== [1] keycloak-0 에서 로그인 ===')
print(f" sid {SID}")
print(f" sub {ac.get('sub')}")
print(f" iss {ac.get('iss')}")
print(f" access 수명 {ac['exp']-ac['iat']}초")
print(f" refresh 수명 {rc['exp']-rc['iat']}초 typ={rc.get('typ')}")
print(f" refresh jti {rc.get('jti')}")
print()
print('=== [3] 같은 sid 가 두 노드 모두에서 보이는가 ===')
for step, who in (('STEP3_SESSIONS_K0', 'keycloak-0 (발급 노드)'),
('STEP3_SESSIONS_K1', 'keycloak-1 (반대편)')):
d = j(get(step))
if d is None:
print(f' {who:24} 파싱 실패: {get(step)[:120]}'); continue
ids = [s.get('id') for s in d]
mark = '보임 ✔' if SID in ids else '없음 ✘'
print(f' {who:24} 세션 {len(ids)}개 중 대상 sid → {mark}')
for s in d:
if s.get('id') == SID:
print(f" ipAddress={s.get('ipAddress')} start={s.get('start')} lastAccess={s.get('lastAccess')}")
def show(step, title, expect):
print(); print(f'=== {title} ===')
body = get(step)
code = [l for l in body.splitlines() if l.startswith('http_code=')]
code = code[0].split('=')[1] if code else '?'
d = j(body)
ok = '기대대로' if code == expect else f'기대({expect})와 다름'
print(f' HTTP {code} ← {ok}')
if d and 'access_token' in d:
c = claims(d['access_token'])
same = '동일 ✔' if c.get('sid') == SID else f"다름 ✘ ({c.get('sid')})"
print(f' 새 토큰의 sid → {same}')
elif d:
print(f" error {d.get('error')}")
print(f" error_description {d.get('error_description')}")
show('STEP5_REFRESH_ON_K1',
'[5] keycloak-0 이 발급한 refresh token 을 keycloak-1 에 사용', '200')
print(); print('=== [6] keycloak-1 을 통해 로그아웃 ===')
print(' ' + get('STEP6_LOGOUT_VIA_K1').strip())
show('STEP7_REFRESH_ON_K0_AFTER_LOGOUT',
'[7] 로그아웃 후 keycloak-0 에서 갱신 시도 (무효화 전파)', '400')
open('/tmp/session-replication/sid.txt','w').write(SID)
PY
SID=$(cat /tmp/session-replication/sid.txt 2>/dev/null)
echo
echo "=== [8] PostgreSQL 에서 그 sid 를 직접 확인 ==="
echo " 대상 sid: $SID"
$PSQL "select user_session_id, offline_flag, created_on, last_session_refresh
from offline_user_session where user_session_id='$SID'" 2>/dev/null \
| sed 's/^/ /' | grep -q . \
&& $PSQL "select user_session_id||' | flag='||offline_flag||' | created='||created_on||' | refresh='||last_session_refresh
from offline_user_session where user_session_id='$SID'" 2>/dev/null | sed 's/^/ /' \
|| echo " 행 없음 — 로그아웃으로 삭제되었다"
echo
echo " 전체 세션 수: $($PSQL 'select count(*) from offline_user_session' 2>/dev/null)"