DAVx5 (CalDAV/CardDAV) reported the exact same 'unexpected end of
stream' / EOFException error again at 2026-08-10T04:00:58Z, roughly
15 minutes after PR #39 deployed. That PR's backend-side fix (option
http-server-close on 'backend next') is confirmed working -
journalctl/nginx access logs show a completely clean, uninterrupted
request stream on the haproxy<->nginx leg through the exact failure
timestamp.
Root cause of the recurrence: PR #39 also added 'timeout
http-keep-alive 30s' to defaults, intended as an unrelated tidy-up
given maxconn=80. That value didn't account for client-side HTTP
connection pooling: DAVx5 runs on OkHttp, which holds idle pooled
connections open for up to 5 minutes by default before evicting them.
With haproxy closing idle client-facing keep-alive connections after
just 30s, any DAVx5 connection idle between 30s-300s got silently
closed by haproxy while the client still considered it live - the
client's next reuse attempt produced exactly the same class of error,
just relocated from the haproxy<->nginx leg to the client<->haproxy
leg instead of being fixed.
Fix:
- defaults: raise 'timeout http-keep-alive' from 30s to 6m, safely
above OkHttp's 300s (5min) idle-eviction default, so a client's own
pool always evicts a stale connection before haproxy would.
- backend next: add 'log-tag next' so this backend's haproxy log
lines carry a distinct syslog tag ('journalctl -t next') instead of
being interleaved with every other backend under the shared
'haproxy' tag - this specific incident took significant manual
grep/awk work to isolate 'next' traffic from git/matrix/immich noise
in the same log stream, which a dedicated tag eliminates going
forward.
Verified by comparing haproxy's own next/nextcloud access log lines
(all showing normal termination, no CD/SD flags) against nginx's
nginx_access journal (clean, continuous, no gap) across the exact
04:00:58 UTC failure window - confirming the backend-side legs were
healthy and the failure had to be on the client<->haproxy leg instead.
Could not run 'haproxy -c' locally (no toolchain in the agent
sandbox) - recommend confirming via CI/garnix before merge, same
caveat as prior PRs in this series (#37, #38, #39).
DAVx5 (CalDAV/CardDAV) on greg's phone was intermittently failing every
sync type (CONTACTS/EVENTS/TASKS/RefreshCollectionsWorker) against
next.thehellings.com with:
java.io.IOException: unexpected end of stream
Caused by: java.io.EOFException: \n not found: limit=0
This is the classic OkHttp/HTTP client signature of the far end
silently closing a pooled keep-alive connection: the client reuses a
socket it still believes is open, gets zero bytes back while reading
response headers, and throws exactly this exception.
Root cause: HAProxy's 'next' backend proxies to nginx on
127.0.0.1:8080, and HAProxy defaults to end-to-end keep-alive (both
client- and server-side) unless told otherwise. nginx's
keepalive_timeout is 65s, so any HAProxy<->nginx connection idle past
that gets closed by nginx without HAProxy's knowledge. A request that
lands on that now-dead pooled connection right after gets nothing back
- surfacing to the client as a bare socket EOF while reading headers.
The frontend's existing 'option http-server-close'/'http-keep-alive'
pair only governs the client-facing side of HAProxy and does nothing
for the HAProxy->nginx leg.
Fix:
- backend next: add 'option http-server-close' so HAProxy opens a
fresh connection to nginx per request instead of pooling/reusing
one. The backend is localhost, so the extra TCP handshake cost is
negligible, and this removes the whole class of stale-connection EOF
errors.
- defaults: add 'timeout http-keep-alive 30s' to bound how long an
idle client-facing keep-alive connection is held open. Previously
unset, it fell back to 'timeout client' (500s) - unnecessarily long
given maxconn is only 80, and tightens client-side connection churn
to be more predictable too.
Diagnosed by pulling the nginx_access journal (enabled in #37/#38) for
the failing sync window and cross-referencing nginx's
services.nginx.appendHttpConfig / generated nginx.conf keepalive
settings against HAProxy's request-level defaults. Could not run
'haproxy -c'/'nginx -t' locally (no toolchain in the agent sandbox) -
recommend confirming via CI/garnix before merge, same as #38.
The nginx_access syslog tag added in #37 (feat/emily-incident-logging)
used tag=nginx-access. nginx's syslog sink only accepts alphanumeric
characters and underscores in the tag field, so the generated
nginx.conf failed its config test on linode:
nginx: [emerg] syslog "tag" only allows alphanumeric characters
and underscore in .../nginx.conf:114
Because nginx-pre-start failed, nginx.service crash-looped until it
hit systemd's start-limit-hit and gave up entirely. Since Nextcloud is
proxied through nginx (127.0.0.1:8080, fronted by haproxy's 'next'
backend), this took next.thehellings.com down with a 503 from haproxy
(phpfpm-nextcloud/postgresql/redis backends were all healthy and
unaffected - purely an nginx config parse failure).
Fix: use an underscore (nginx_access) instead of a hyphen.
Triggered by investigating a several-hour >10Mbps traffic spike to
linode. HAProxy's own IPAccounting confirmed ~121GB moved over ~19.6h
before it crash-looped, but with 'option httplog' commented out and no
per-backend request logs, there was no way to attribute that traffic
to a specific backend, host, or client.
- linode: enable HAProxy httplog + defaults 'log global' (was
commented out) so every proxied HTTP request is now logged with
timing/status/bytes.
- linode: add a haproxy 'stats' listener on 127.0.0.1:8404 for live
per-backend/per-server connection and byte counters.
- linode: route nginx (Nextcloud's local vhost) access logs to
journald via syslog, since the read-only monitoring account has no
access to /var/log/nginx/*.
- linode: enable vnstat for historical per-interface bandwidth
tracking (5-min granularity) so a reported 'traffic was high for N
hours' can be confirmed/timestamped immediately instead of
reconstructed after the fact from journal timestamps.
- k3s manifests: enable Traefik access logging (JSON) — this is the
ingress layer HAProxy forwards :80 traffic to (git/matrix/immich),
and lacked any per-request visibility.
- hosts/baseline.nix (fleet-wide): add a journald rate limit
(2000 lines / 30s per unit). Found live while investigating that
uptime-kuma on 'kuma' was logging a Prometheus label-validation
error on every monitor beat (~100k lines/hour), which was itself
degrading journalctl responsiveness on that host during the
cross-host traffic scan.
Related but not otherwise addressed here: Nebula relay/handshake
churn on kuma's tunnel and the etcd read-latency warnings seen on
isaiah/zeke around the same incident window — noted for a future
investigation, not fixed by this PR.
Adds a new NixOS module (greg.monitoring-access) that provisions a
dedicated, SSH-key-only 'emily' user account across all managed hosts.
The account is intentionally minimal-privilege:
- No password set (SSH key auth only)
- Not a member of wheel, no sudo/sudo-rs rules
- Only extra group membership is systemd-journal, granting read access
to system logs for monitoring/analysis tasks
- Authorized key lives in home/ssh/emily_authorized_keys, mirroring the
existing pattern used for the greg account's authorized_keys
This lets the Hermes agent (emily) log in read-only to inspect logs and
system state when asked, without any ability to modify configuration,
escalate privileges, or run destructive commands.
Module is imported unconditionally in modules/nixos/default.nix like
the other nixos modules, and defaults to enabled; it can be disabled
per-host via greg.monitoring-access.enable = false if ever needed.
- Remove builder2 (retired host): dangling network.json entry and
empty home/hosts/builder2 stub. Its old IP (10.42.1.17) is already
correctly owned by pve4.
- Remove the abandoned Ceph module (modules/nixos/ceph.nix) and its
unencrypted plaintext keyring files under secrets/. No host ever
enabled services.ceph-benaco; the keyrings were dead, unencrypted
credentials sitting in the repo.
- Normalize Darwin host identity: IVR and Lithic are the only two
physical Darwin machines. All other darwin/hosts/* names are now
symlinks to whichever of the two they represent, matching the DHCP
name variations nix-darwin sees depending on network:
gregory -> ivr
gregory.hellings-mbp -> ivr
MacBook-Prolocal -> ivr
gregs-MacBook-Pro-16-inch-Nov-2024 -> lithic
li -> lithic (pre-existing)
This lets each machine's config be maintained once regardless of
what hostname it currently advertises.
pve1 is a static/DHCP-reserved Proxmox host at 10.42.0.4 (previously
mislabeled 'joel' in some places). 10.42.1.1 is the UDM Pro gateway
IP, not pve1 -- OPNsense was retired in favor of Ubiquiti. Removes
the stale duplicate PVE1 DHCP reservation at 10.42.1.1 and drops the
now-redundant 'joel' entry from network.json (consolidated into
pve1).
* keepalived does not work with Nebula VPN
* update Genesis firewall to allow passing through local traffic
* target all traffic directly to the LAN IP using genesis's routing
* Consolidate Linode into a single file
* Convert gitea and matrix to using Nebula connections
* Have Linode proxy to Nebula connections instead of Tailscale
* Update Acme to use DNS-01
* Update Flake to pull from branch that supports ACME 5.x client
Zims update script has been slightly mangled since nix-prefetch stopped
working.
Now it is updated to use nix-prefetch-url and no longer pulls from the
Torrent sources. That script exports a regular SHA256 hash and not an
SRI signature, so we now convert that to SRI as a second step in the
pre-fetch pipline
Also adding a cron to run the tool every month on the first, in order to
keep it up to date.
greg
emily
Greg Hellings