Container Apps Sizing
Review and act on sizing recommendations for container apps and jobs on the Consumption profile, and for Dedicated workload profiles. Open Recommendations > Sizing and set the Resource type filter to Container Apps or Dedicated Workload Profiles. Savings are USD list price, before Azure's monthly free grant.
9 Sections
Two resource types
- Container Apps
- Container apps and container app jobs on the Consumption profile. One recommendation per app or job.
- Dedicated Workload Profiles
- The Dedicated D and E workload profiles of a Container Apps environment. One recommendation per environment, listing each sized profile.
What's Analyzed
Container Apps sizing covers single-container apps and jobs on the Consumption profile, using up to 90 days of Azure Monitor platform metrics collected by the scan. There's nothing to enable in Azure. Cost data improves the accuracy of app savings and is required for scale to zero.
Not analyzed
- Azure Functions on Container Apps and Logic Apps (the platform generates their scale rules), and Arc-connected environments.
- Apps in multiple revision mode, apps with init containers, and multi-container apps.
- Jobs that run more than one replica or execution at a time. Event-driven jobs are analyzed only when their maximum executions is set to 1 (the Azure default is 100).
- Apps and jobs that are stopped, being provisioned, or in a Degraded or Failed state.
- Apps on a Dedicated profile. Those are covered under Dedicated Workload Profiles.
- Allocations that aren't a standard Consumption size.
Apps with a CPU or memory utilization scale rule get no allocation decrease, because the scaler reacts to a percentage of the allocation and a smaller one would change scaling. They can still get a Lower minimum replicas recommendation.
Toggles
Toggle Options has a Container Apps section with one switch, Lower minimum replicas, and a Dedicated Workload Profiles section with three: Smaller profile size, Profile family change and Move to Consumption. With several on, the page shows the eligible option with the largest saving.
Every switch is off by default
The default view shows only the allocation decrease for apps and jobs, and lower minimum nodes for Dedicated profiles.
The Recommendation type filter offers Size down and Lower minimum replicas for Container Apps, and Lower minimum nodes, Smaller profile size, Profile family change and Move to Consumption for Dedicated Workload Profiles.
How a Target Is Chosen
Targets are always a standard Consumption size: steps of 0.25 vCPU with 2 Gi of memory per vCPU, from 0.25 vCPU / 0.5 Gi up to 4 vCPU / 8 Gi. Consumption-only environments (no workload profiles) stop at 2 vCPU / 4 Gi. Every target must pass these checks:
- Memory: the highest daily memory peak on any running day must stay under the memory ceiling share of the new limit (default 60% for apps, 50% for jobs).
- CPU: the P95 of the daily CPU maximum stays under the aggressiveness target (Conservative 70%, Balanced 80%, Aggressive 90%), and the window's CPU peak fits in the new allocation.
- One step at a time: the target keeps at least half of the current CPU and memory.
- Savings: at least $1 a month.
Memory is a hard limit
Memory is checked against the highest peak in the window, not a P95. Aggressiveness only changes the CPU target.
Larger cuts come in steps
After you apply a recommendation and it runs for a while, a later scan can recommend a further step down.
The window is up to 90 days, and only running days count: days an app had at least one replica, or a job used memory. There's no recommendation at all when, in the window, a running day had a replica restart, a job had a failed execution, memory grew (the last 7 running days peaked more than 25% above the earlier peak), a running day is missing its memory value, or the newest metrics are more than 14 days old.
Lower minimum replicas
Apps only, behind the Lower minimum replicas switch. It's offered when the app has a minimum of 1 or more, a scale rule can add replicas back (an app with ingress and no rule has Azure's default HTTP rule), and the maximum replicas is above the target. The target halves the current minimum, rounded down and never below 1, or goes to 0 when scale to zero is safe. The saving comes from the replicas held only by the minimum, which bill at the idle rate.
When the remaining replicas can carry the load on a smaller allocation, the option combines the allocation decrease with the lower minimum. Otherwise it lowers the minimum at the current allocation.
Scale to zero needs all of these
- No CPU or memory scale rule, and no Dapr.
- An HTTP, TCP or custom (non CPU/memory) scale rule.
- Ingress or a custom rule that can start the app again.
- Cost data showing the replicas held by the minimum were billed at the idle rate at least 90% of the time. Busy request-driven apps therefore stay at a minimum of 1.
How savings are priced
- Savings are based on usage over the observed span: replica time for apps, run time for jobs.
- App rates blend the active and idle rates from the app's last 30 days of cost data. Without cost data the app is priced at the idle rate, which understates the saving; the card says so.
- Jobs are always priced at the active rate. Job run time is estimated from metrics and tends to be lower than the bill, so job savings lean low.
- Usage is overstated for an app stopped for part of the window, or whose minimum went from 0 to 1 or more inside it.
Savings are before the monthly free grant
The free grant is per subscription and can't be assigned to one app, so small apps can show a saving the free grant would have absorbed.
Dedicated Workload Profiles
StratoLens sizes the D and E profiles that run at least one app or job. For each day, it works out how many nodes the profile needed by packing that day's replicas onto nodes by their CPU and memory allocations, the way Azure places replicas, keeping 15% of each node's memory free. Because node counts come from allocations, there's no aggressiveness setting.
- Sizes considered: D4, D8, D16 and D32 (general purpose), E4, E8, E16 and E32 (memory optimized). A size is considered only when every replica fits on one node.
- Skipped: GPU and Flexible profiles, and any profile with a workload in multiple revision mode, a job that runs in parallel, an allocation StratoLens can't read, or a workload with no metrics.
- Empty profiles are listed under Orphaned Resources as Empty Dedicated Workload Profile, not here.
Levers
- Lower minimum nodes
- Always on, no switch. Same size, with the minimum set to the P95 of daily required nodes when that is below the current minimum.
- Smaller profile size
- A smaller size in the same family, for example D8 to D4.
- Profile family change
- Any size of the other family, D to E or E to D.
- Move to Consumption
- Moves every app and job on the profile to Consumption, each at the smallest Consumption size that covers its allocation, priced at the active rate. Always an estimate.
Per profile, the option with the highest saving wins. On a tie, Lower minimum nodes wins, because it changes nothing in place. Savings are USD list price from node counts, not the bill.
The Plan Management fee is never a saving
The environment's Plan Management fee is part of the current cost on both sides of the card, even when every Dedicated profile would leave. A private endpoint or planned maintenance keeps the fee.
Reading the Card
Container Apps card
The title is Allocation Size Down Recommendation, or Lower Minimum Replicas Recommendation when the chosen option lowers the minimum replicas. An App or Job badge shows the resource kind, and an option badge shows Size Change, Lower minimum replicas, or both joined by "+".
- Current → Recommended
- Allocation and monthly cost on each side, for example "0.5 vCPU / 1 Gi". Minimum replicas appear, as in "0.5 vCPU / 1 Gi, min 1 replica", only when the option lowers them.
- Warnings
- Caveats for the chosen option. Always present.
- Utilization Evidence
- Memory (peak per replica against the new limit, for example "Peak 245 MiB → 0.5 Gi limit (48%)"), Memory ceiling, CPU (daily max) as P95 and peak in cores, and Running days. Values are absolute, per replica.
- "Priced at the idle rate: no cost data."
- StratoLens had no cost data for the app, so the saving is understated.
Dedicated Workload Profiles card
The title names the most disruptive change in the chosen option: Lower Minimum Nodes Recommendation, Smaller Profile Size Recommendation, Profile Family Change Recommendation or Move to Consumption Recommendation.
- Current → Recommended
- Dedicated profile count before and after (for example "2 Dedicated profiles" → "1 Dedicated profile"), with cost and the Plan Management fee on both sides.
- Profile rows
- One per sized profile: Saves $N/mo or No change, current (for example "D4 × min 2 (avg 2.3 nodes)") to target ("D4 × min 1", "E4 × min 1" or "Consumption"), the number of apps and jobs, Required N nodes (P95), Observed avg, days of node data, and the row's own warnings. Row savings add up to the card total.
- "Required: not enough data"
- The profile has fewer days of node data than the minimum data window and shows No change. Other profiles in the environment are still sized.
On both cards
- An Environment line names the Container Apps environment.
- A commitment banner appears when a reservation or savings plan covers the app, job or environment.
- Estimated Savings per month, per year and % reduction, and Confidence Analysis with the same three factors as other sizing types. Container Apps confidence is capped at Medium below 30 running days.
- Severity uses the default view with every switch off: above $200/month is High, above $50 is Medium, otherwise Low. An app whose only option needs Lower minimum replicas is Low, however large the saving.
View Performance opens the app or job in Performance Metrics; for a Dedicated recommendation it opens the environment's node count per profile. View in Explorer, Hide and Unhide work as on other resource types.
Before You Apply a Change
- Apps: changing CPU or memory creates a new revision. In single revision mode the old revision keeps running until the new one is ready.
- Jobs: every later execution uses the new allocation.
- Smaller memory limit: watch for replica restarts after the change.
- Java apps: automatic memory fitting turns off below 1 GB. Check the heap settings before going below 1 Gi.
- Lower minimum nodes: new replicas can wait for a node after a scale-out.
- Move to Consumption: apps and jobs lose single-tenant isolation, and Consumption replicas are limited to 8 GB of images.
- Reservations or savings plans: resizing leaves reserved capacity unused unless other eligible usage in the region uses it.
Scale to zero changes how the app behaves
The first request after an idle period has a cold start, and in-process background work such as timers, queue polling and cache warmers stops while the app is at zero.
Dedicated size and family can't change in place
Add a new profile, move each app and job to it, then delete the old profile. Moving an app or job keeps its revision but replaces every replica, which is a restart.
After a Change
- The recommendation leaves the active list on the next scan, and the change appears under Show Resolved with the saving credited. Credited savings are estimates at USD list price, because the usage history still holds mostly pre-change days.
- While an app or job is still provisioning, or is Degraded or Failed after the change, StratoLens holds the recommendation for up to 3 days. A change that broke the app is never credited.
- After a minimum-replicas change, usage figures take up to 90 days to settle to the new replica count.
- A Dedicated size or family change shows as "new profile added, old profile removed". A change still in progress (an old profile emptied but not yet deleted while its minimum is above 0) is held for up to 3 days.
- If you delete a Dedicated profile that was also flagged under Orphaned Resources, both features record the saving.
Settings
Open Settings > Scanner, expand Sizing Recommendations and find the Container Apps and Dedicated Workload Profiles blocks. Like other scanner settings, changes apply on the next scan, and saving needs the Configure Scanner Settings permission. Viewing the Sizing page needs only data read access.
Container Apps defaults
- Sizing Aggressiveness
- Balanced (80%). Sets the CPU target only.
- Minimum Data Window (days)
- 30, range 0 to 90. Apps and jobs with fewer running days get no new recommendations; existing ones are kept. Below 7 lowers the statistics trust floor, and 0 is testing mode.
- Memory ceiling, apps (%)
- 60, range 30 to 80. The highest observed memory must stay under this share of the new limit. Lower is safer.
- Memory ceiling, jobs (%)
- 50, range 30 to 80. The same rule, for jobs.
Dedicated Workload Profiles defaults
- Minimum Data Window (days)
- 30, range 0 to 90. Profiles with fewer days of node count data get no sizing target. There's no aggressiveness setting.
The first scan after install, or after upgrading to the version with Container Apps sizing, collects up to 93 days of history (what Azure keeps), so apps with 30 or more running days can get a recommendation right away. Recommendations also need Container Apps price data for the region; a region without it is skipped until the daily SKU Catalog Update has run.
Troubleshooting
An app or job I expected has no recommendation
Answer
Check, in order: it's on a Dedicated profile (look under Dedicated Workload Profiles); it's a type that isn't analyzed; it's already at 0.25 vCPU / 0.5 Gi; it had a restart or failed execution in the last 90 days; memory has grown recently; it has fewer running days than the minimum data window; the newest metrics are over 14 days old; the region has no price data yet; the saving is under $1/month; or it's hidden.
My app has a CPU scale rule and only shows Lower minimum replicas
Answer
This is expected. Apps with a CPU or memory utilization scale rule get no allocation decrease.
I turned on Lower minimum replicas but nothing changed
Answer
The app needs a minimum of 1 or more, a scale rule (or ingress), and a maximum above the target. Jobs never get this option.
Scale to zero is never offered
Answer
See the scale-to-zero conditions. Common blockers are a CPU or memory scale rule, Dapr, no cost data, or an app busy enough that its minimum replicas aren't mostly idle.
My weekly job never gets a recommendation
Answer
A weekly job has about 13 running days in the 90-day window, below the 30-day minimum data window. Lower the Container Apps minimum data window to cover it.
An empty Dedicated profile isn't listed here
Answer
Empty profiles are listed under Orphaned Resources as Empty Dedicated Workload Profile.
The savings don't match my bill
Answer
Savings are USD list price, before the monthly free grant and any discounts or commitments. Job run time and Move to Consumption costs are estimates.
Severity is Low even though the saving is large
Answer
Only savings available with every switch off count toward severity.
My recommendation disappeared after I resized
Answer
This is expected. The change moves to Show Resolved after the next scan.