Azure App Service is a service used to create and deploy scalable, mission-critical web apps.
The roughly 25-hour LinuxContainerRecycle cadence is consistent with App Service on Linux’s periodic platform recycle, but the available customer-visible telemetry cannot identify the specific system process behind the preceding memory spike or prove that a memory/swap threshold caused the worker replacement.^1^
What the evidence supports
- Daily spike: App Service runs platform processes for security updates, Kudu/SCM, monitoring, authentication, and other features. Their memory is included at plan level even when application activity is low.^2^ However, public diagnostics do not expose enough host-level process information to name which platform process produced this particular spike.
- Worker change: During maintenance, App Service can replace an assigned worker with an updated worker. That produces a restart on a different machine without requiring a deployment or customer-initiated Activity Log event.^3^
- Memory limit: A graph reaching 95–99% does not by itself prove enforcement of a hard memory or swap threshold. On Linux, high memory utilization alone is not equivalent to memory pressure; stronger indicators are low available memory, swapping, or an OOM-kill event.^4^ The documented “Percent Memory limit” recycle applies to 32-bit worker processes, whereas Linux App Service runs 64-bit by default, so it should not be applied to this case without that exact detector message.^5^
- The second B1 plan surviving a similar spike also argues against treating the displayed percentage alone as a deterministic replacement threshold.
The ordering remains important: because memory rises before the recorded move, it cannot safely be described as merely post-recycle warm-up. It could be platform work associated with the recycle, host memory pressure, or another worker-level event; only Microsoft’s internal host telemetry can distinguish those possibilities.
Capture the next occurrence
For each app, open Azure portal → App Service → Diagnose and solve problems and preserve results around the event from:
- Availability and Performance → Memory Usage
- Linux Swap Space Low
- Process List and Process Fill List
- Web App Restarted
- Search for Web App Slow, then inspect:
- Worker Process Events
- Instance Allocation Events
- Platform Observations
Those detectors are supported for Linux and can distinguish app-process growth, swap pressure, container recycling, VM failover, and update-domain movement.^6^^7^
Record the following for escalation:
- UTC start and end of the memory spike.
- Old and new worker hostnames.
-
LinuxContainerRecycleandSiteStartedtimestamps. - Plan-level memory and swap results.
- Per-app/process memory immediately before the gap.
- HTTP failure interval and startup/warm-up duration.
- Whether the detector reports an allocation event, maintenance, OOM, or Auto-Heal action.
Also inspect Diagnose and solve problems → Auto-Heal. Auto-Heal supports memory-based rules on Linux; if no rule is configured, do not enable one merely to address a platform worker move because it can add another restart.^8^
Preventing the downtime
There is no documented control on a multitenant B1 plan that pins the app to one worker, disables platform recycling, or selects its maintenance window. Maintenance timing control is limited to App Service Environment v3.^3^
For supported availability mitigation:
- Scale out to at least two active instances.
- Add a lightweight endpoint that returns
200only when startup dependencies are ready. - Configure it under App Service → Health check.
- Set
WEBSITE_WARMUP_PATHto an appropriate initialization endpoint if startup readiness is contributing to the 5–22 minute interruption.
Health Check on one instance only monitors it; the load balancer cannot remove the sole unhealthy instance without taking the application down. Its traffic-rerouting benefit requires at least two active instances and sufficient plan capacity.^9^
Because 5–22 minutes is substantially longer than a simple container restart, submit an Azure support case if the detectors do not identify the trigger. Include several exact UTC windows and request correlation of host memory/swap pressure, worker-reallocation reason, and platform process consumption. That internal telemetry is required to identify the process and determine whether memory pressure initiated the replacement.
References
- FAQ - Azure App Service on Linux - Azure | Microsoft Learn
- Web apps performance, availability, and application issues for Azure App Service
- Routine maintenance for Azure App Service, restarts, and downtime
- Troubleshoot performance issues and isolate bottlenecks in Linux
- "Worker Process requested recycle due to 'Percent Memory' limit" message
- Diagnostics in Azure App Service
- Troubleshoot slow web app performance in Azure App Service
- Best practices for Azure App Service
- Monitor App Service instances by using Health check