Your dashboard shows requests per second falling off a cliff, but nobody changed anything, the load balancer still reports healthy demand, and there is no error spike to point at. The most natural reading, “traffic went away,” is usually wrong.
The trap is in how the number is produced. The request rate most operators watch comes from mod_status Total Accesses, which counts completed requests. It is not a measure of arriving demand. When Apache cannot complete requests, because every worker is stuck or a backend has stopped answering, the completed-request rate collapses even while clients are still hammering the front door. The server has not lost its traffic. It is silently shedding load, and the metric you trust is the last place that shows up.
This guide covers why the counter lies during saturation, how to confirm which of the four usual causes you are dealing with, and what to do about each one.
What this means
A request travels a short path before it shows up in your RPS graph: the kernel accepts the TCP connection into the listen backlog, an Apache worker takes it off the queue, the request is processed (possibly via a backend), the response is sent, and only then is the request counted in Total Accesses and written to the access log.
If anything downstream of “worker takes the connection” stalls, requests pile up in two invisible places: the kernel listen backlog and the clients themselves. Neither increments your request counter. So you get the signature pattern of silent load shedding:
Total Accessesrate drops toward zero.- The access log develops a gap, because the access log only records completed requests.
- Clients experience timeouts or connection resets that your metrics do not show.
- CPU often looks calm, because waiting is free.
The four causes that produce this pattern are worker exhaustion (all slots busy on something slow), a hung backend holding every proxied request open, an upstream change (load balancer or DNS) that actually did divert traffic away, and, less commonly, a local resource wall like file descriptor exhaustion that prevents Apache from accepting at all. Distinguishing them fast matters, because the fixes are completely different.
flowchart LR client[Clients] --> lb[Load balancer] lb --> backlog[Kernel listen backlog] backlog --> workers[Apache workers] workers --> backend[Backend] backend --> done[Request completed] done --> counted[Total Accesses + access log] backlog -. "queued, invisible" .-> gap[RPS drop] workers -. "stuck in W state" .-> gap
Common causes
| Cause | What it looks like | First thing to check |
|---|---|---|
| Worker exhaustion | BusyWorkers at MaxRequestWorkers, IdleWorkers zero, Recv-Q growing, AH00484 in error log | server-status BusyWorkers and IdleWorkers |
| Hung backend (proxy deployments) | Scoreboard full of W states, normal CPU, 5xx trickling in, request rate paradoxically low | Curl the backend directly, bypassing Apache |
| Upstream routing change | RPS drops but workers are idle, backlog empty, no errors anywhere | LB target health and DNS resolution for your hostname |
| Local resource wall (FDs, disk) | Intermittent 5xx, “Too many open files” or “No space left” in error log, scoreboard stuck in L | Error log tail and df -h on the log filesystem |
Quick checks
All of these are read-only and safe to run during an incident.
# Worker utilization: the single most important first look
curl -s http://localhost/server-status?auto | grep -E "BusyWorkers|IdleWorkers"
# Scoreboard state distribution: where are workers stuck?
curl -s http://localhost/server-status?auto | grep "Scoreboard:" | \
awk '{print $2}' | fold -w1 | sort | uniq -c | sort -nr
# Listen backlog depth: Recv-Q = queued connections, Send-Q = limit
ss -ltn | grep -E ':80\s|:443\s'
# MaxRequestWorkers reached events (definitive exhaustion signal)
grep "AH00484" /var/log/apache2/error.log | tail -20
# Access log gap tell: is the log still being written at all?
tail -5 /var/log/apache2/access.log; date
# Error log: FD exhaustion, disk full, proxy failures
grep -E "\[error\]|\[crit\]|\[alert\]" /var/log/apache2/error.log | tail -20
# Backend direct probe (adjust host:port), bypassing Apache entirely
curl -s -o /dev/null -w "TTFB: %{time_starttransfer}s Total: %{time_total}s\n" \
--max-time 10 http://backend-host:8080/health
# Established connections to your backend, if proxying
ss -tn state established dport = :8080 | wc -l
On RHEL-family systems the logs live under /var/log/httpd/ (error_log, access_log) instead of /var/log/apache2/.
How to diagnose it
Establish the baseline first. Compare the current RPS to the same time of day from previous days, not to an absolute number. A drop from 500 to 50 req/s at 03:00 may be normal; the same drop at 14:00 is an incident. If you have no time-of-day baseline, treat any sudden halving as suspect until proven otherwise.
Check worker utilization. If
BusyWorkersequals your configuredMaxRequestWorkersandIdleWorkersis zero, Apache is saturated and queuing. Requests are arriving but not completing, so your RPS drop is load shedding, not lost demand.MaxRequestWorkersis not shown in server-status; read it from the config, and rememberServerLimitcan silently cap it.Read the scoreboard states. Saturation alone does not tell you why. A scoreboard dominated by
W(sending reply) with calm CPU and memory means workers are waiting on something, almost always a backend in proxy deployments. ManyR(reading) states points at slow clients or a Slowloris pattern. ManyL(logging) states means the log pipe or disk is the stall.Confirm with the access-log gap tell. If the access log has stopped advancing while the error log is quiet, Apache is alive but unable to complete requests. Queued connections never reach a worker and are never logged. A gap in the access log concurrent with an AH00484 event is expected behavior, not a logging bug.
Check the listen backlog. A sustained non-zero
Recv-Qon ports 80/443 means connections are arriving faster than Apache accepts them. This is the earliest saturation signal and confirms real arriving demand even though RPS has collapsed. The kernel caps the effective backlog atnet.core.somaxconnregardless of theListenBacklogdirective.If workers are idle, look upstream. Low BusyWorkers, empty backlog, and no errors means the traffic genuinely is not arriving. Check the load balancer: has this node been removed from rotation by a failing health check? Did a DNS change or expiry repoint the hostname? Did a WAF or CDN rule start blocking? This is the one cause where “traffic went away” is actually true, and no amount of Apache tuning will bring it back.
If proxying, test the backend directly. A hung backend holds workers in
WuntilProxyTimeoutfires, so requests trickle out as 5xx or never complete at all. Curl the backend from the Apache host, bypassing the proxy. If it hangs, your problem is the backend and Apache is the victim.Rule out local resource walls. “Too many open files” in the error log means FD exhaustion; “No space left on device” plus a scoreboard full of
Lstates is the log-stall deadlock. Both stop completions while the process looks alive.
Metrics and signals to monitor
| Signal | Why it matters | Warning sign |
|---|---|---|
| BusyWorkers / MaxRequestWorkers | Primary saturation gauge; the drop to “zero RPS” is nearly instantaneous at 100% | Sustained above 80%, IdleWorkers at zero |
| Scoreboard state distribution | Tells you what workers are stuck on, not just that they are stuck | W dominating with calm CPU; any L pile-up; R above ~20% |
| Listen backlog Recv-Q | Leading indicator that arrives before user-visible failure | Sustained non-zero during normal traffic |
| AH00484 events | Apache explicitly reporting worker exhaustion | Any occurrence |
| Access log write rate | The gap tell: completions stopped even if the process is alive | Log stops advancing while error log is quiet |
| Backend response time | The number one cause of worker holding in proxy setups | P95 above 2x baseline |
| 503/504 rate | Backpressure surfacing as errors after the silent phase | Any sustained occurrences |
Fixes
Workers saturated by a hung backend
This is the most common root cause and the most commonly mistreated one. Raising MaxRequestWorkers does not help; it just gives the backend more workers to hang.
- Take the node out of LB rotation if the backend is dead, and fix the backend.
- If the backend is slow but alive, temporarily lower
ProxyTimeoutso Apache fails fast and frees workers instead of holding them for the full default. Tradeoff: clients see quick 5xx errors instead of long hangs, which is usually the better failure mode but still an error. - Longer term, size the proxy connection pool deliberately. The default
maxequalsThreadsPerChild(1 on prefork), which is far too small for production and produces 503s under moderate load that look exactly like Apache saturation.
Workers saturated by real demand
If the scoreboard shows healthy turnover, backends are fast, and demand simply exceeds capacity:
- Raise
MaxRequestWorkers, but only after the memory math:MaxRequestWorkers x per-child RSSmust stay under roughly 70% of RAM. Setting it from a round number without measuring per-process memory is how you trade a queuing incident for an OOM incident. - On prefork or worker MPM, check
KeepAliveTimeout. High values hold worker slots on idle keepalive connections; event MPM largely removes this concern by handling keepalive connections asynchronously instead of occupying a worker. - If the scoreboard shows many
Rstates instead, treat it as slow clients or Slowloris: confirm source IP concentration and make suremod_reqtimeoutis active, then block offenders at the firewall, not in Apache.
Upstream routing change
- Fix it where it lives: LB target group health, DNS records, CDN or WAF rules.
- If the LB removed the node because its health check times out during saturation, that is worker exhaustion feeding back into routing. Fix the saturation first; the LB will re-add the node.
- Verify the LB health check tests the real service path. A health check that fetches a static file passes while the proxied application is dead, which is how nodes stay in rotation while serving nothing.
Local resource walls
- FD exhaustion: raise
LimitNOFILEin the systemd unit (this needs a restart, not a graceful reload), and consolidate per-VHost log files if you have many virtual hosts. - Log stall: free space on the log filesystem immediately, then restart Apache after the disk is fixed so blocked workers clear. Keep logs on their own filesystem so a log explosion cannot take the server with it.
Prevention
- Alert on the ratio, not the rate. RPS drops are a lagging, ambiguous signal. Alert when
BusyWorkers / MaxRequestWorkersexceeds 80% sustained, and page when IdleWorkers is zero with a growing backlog or AH00484 events. That catches silent shedding minutes before the RPS graph flatlines. - Watch the backlog. Sustained non-zero
Recv-Qis the earliest warning that acceptance cannot keep up. - Monitor backends separately. In proxy deployments, backend response time is the single highest-value signal for predicting this exact incident.
- Set
MaxConnectionsPerChildto a finite value (5000-10000) so a leaky module cannot slowly grow children into memory exhaustion, which is another path to the same symptom. - Baseline by time of day. “Dropped to near zero” is only meaningful against the same-hour history.
How Netdata helps
- The Apache collector polls
server-status?autoevery second, soTotal Accessesbecomes a true per-second completed-request rate rather than a lifetime average like theReqPerSecfield, making drops visible within seconds. - BusyWorkers and IdleWorkers are charted alongside the request rate, so you can see the telltale divergence, rate falling while workers pin at maximum, on one screen instead of correlating by hand.
- The scoreboard state breakdown over time shows whether a drop came with a
Wpile-up (backend),Rpile-up (slow clients), orLpile-up (log stall), which is the key diagnostic branch. - Correlating Apache worker saturation with backend latency, system CPU, and socket statistics (listen backlog, connection states) in one place is what turns a 30-minute guessing session into a two-minute root cause.
- Anomaly detection on the request rate flags deviations from the learned time-of-day pattern, which is exactly the baseline comparison this symptom demands.
Netdata’s Apache HTTP Server monitoring with Netdata brings these signals together with per-second metrics and ML anomaly detection.
Related guides
- Apache 500 Internal Server Error: modules, handlers, and misconfiguration
- Apache 502 Bad Gateway: a backend that returned an invalid response
- Apache 503 Service Unavailable: worker exhaustion versus proxy pool exhaustion
- Apache 504 Gateway Timeout: slow backends, ProxyTimeout, and worker pile-up
- Apache 5xx error rate: 500 vs 502 vs 503 vs 504 and what each one means
- Apache AH00558: Could not reliably determine the server’s fully qualified domain name
- Apache backend response time: telling ‘Apache is slow’ from ’the backend is slow’
- Apache balancer member in error state: reading balancer-manager and failover
- Apache BusyWorkers and IdleWorkers: reading worker utilization from mod_status
- Apache CLOSE_WAIT and TIME_WAIT: connection leaks versus normal churn
- Apache error log monitoring: severity levels, AH codes, and what to alert on
- How Apache HTTPD actually works in production: a mental model for operators






