viktor/infra
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

claude-memory: HA (replicas 2 + PDB) to stop recurring MCP disconnects
All checks were successful
ci/woodpecker/push/default Pipeline was successful
Details

Browse source 
The claude-memory MCP backend ran as a single replica with no PDB, so every
voluntary disruption took it to zero for ~30-90s — which surfaced as the
memory MCP "keeps getting disconnected" problem. Disruption sources hitting
the lone pod: the descheduler (every-5-min CronJob, LowNodeUtilization —
caught evicting it live), Keel image bumps, Reloader restarts on the 7-day
DB-password rotation, node drains, and CI deploys.

The local stdio MCP subprocess itself was proven healthy (fast non-blocking
startup, stderr suppressed, graceful degradation), so the fault was purely
backend availability, not the MCP plumbing.

Fix: run 2 replicas (the backend is stateless FastAPI over shared CNPG
Postgres and already has hostname anti-affinity) + restore the PDB at
minAvailable=1 (safe now — the drain deadlock that justified removing it
only existed at 1 replica) + descheduler evict=false to stop the needless
5-min churn. All five disruption sources become zero-downtime rolling events.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Viktor Barzin 2026-06-18 09:13:36 +00:00
parent 48b7be3b14
commit 8787d361dc
1 changed files with 27 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

30
stacks/claude-memory/main.tf
View file

@ -150,7 +150,11 @@ resource "kubernetes_deployment" "claude-memory" {
}
}
spec {
replicas = 1
# 2 replicas (stateless FastAPI over shared CNPG PG) so node drains, Keel
# bumps, Reloader restarts (7d DB rotation), CI deploys, and descheduler
# evictions become zero-downtime rolling events instead of hard outages —
# the latter were surfacing as recurring memory-MCP "disconnects".
replicas = 2
selector {
match_labels = {
app = "claude-memory"
@ -163,6 +167,10 @@ resource "kubernetes_deployment" "claude-memory" {
}
annotations = {
"dependency.kyverno.io/wait-for" = "postgresql.dbaas:5432"
# Skip descheduler eviction — it bounced this pod every ~5min
# (LowNodeUtilization). The PDB below keeps drains/Keel/CI safe at
# 2 replicas; this just stops the needless churn of the MCP backend.
"descheduler.alpha.kubernetes.io/evict" = "false"
}
}
spec {
@ -261,8 +269,24 @@ resource "kubernetes_deployment" "claude-memory" {
}
}
# PDB removed — single replica with minAvailable=1 blocks all node drains.
# claude-memory is non-critical and recovers quickly after rescheduling.
# PDB restored alongside replicas=2 (2026-06-18). The old reason for removing it
# — a 1-replica minAvailable=1 PDB deadlocks node drains — no longer applies at
# 2 replicas: minAvailable=1 lets one pod be drained/evicted while the other
# serves, so voluntary disruptions never take the MCP backend to zero.
resource "kubernetes_pod_disruption_budget_v1" "claude-memory" {
metadata {
name = "claude-memory"
namespace = kubernetes_namespace.claude-memory.metadata[0].name
}
spec {
min_available = "1"
selector {
match_labels = {
app = "claude-memory"
}
}
}
}
resource "kubernetes_service" "claude-memory" {
metadata {