How to Fix Azure Database for PostgreSQL Stuck in Starting State When It Keeps Recurring

Second time in a month. The portal is frozen. The CLI returns SeverBusyWithOtherOperation on every command. This guide covers the immediate recovery steps, the support escalation template, and the post-incident hardening that stops this from happening a third time.

2× / month

Recurrence within 30 days means a persistent underlying condition — not bad luck

Can't fix it yourself

SeverBusyWithOtherOperation on Stop/Start means a backend lock only Microsoft can clear

7 days

Azure auto-starts a stopped Flexible Server after 7 days — this is a known recurrence trigger

Root cause

Must be investigated and documented in the support ticket — a one-time fix without root cause = a third occurrence

Why It Keeps Recurring — This Is Not Normal Behaviour

Asingle occurrence of a PostgreSQL Flexible Server stuck in the Starting state can be written off as a platform transient. The second occurrence within a month on the same server is a signal: something about this server's configuration, workload pattern, or infrastructure placement is making it structurally prone to this class of failure. A support ticket that just asks to "unblock the server" without requesting root cause investigation will result in a third and fourth occurrence.

The SeverBusyWithOtherOperation error means Azure's control plane holds an exclusive lock on the server resource — an internal workflow that never completed cleanly is still marked as running. This lock prevents all new management operations from being accepted. The error text is the same whether the stuck operation was a start, stop, scale, patch, or maintenance event. You cannot distinguish which one caused it from the outside, and you cannot clear it from the portal or CLI.

What makes recurring cases different from a single occurrence is that recurring cases have an identifiable underlying condition that survived the first support fix. The most common recurring triggers are: an active maintenance window that collides with a scheduled stop/start cycle at a consistent weekly cadence; storage approaching capacity that causes WAL accumulation; a burstable (B-series) SKU that repeatedly exhausts CPU credits during startup; or inactive logical replication slots that steadily consume disk space between incidents. None of these clear themselves when Microsoft support force-resets the server state — they must be explicitly addressed.

Figure 1 — How a recurring stuck state forms: the underlying condition persists through the first support fix

Clearing the stuck state without identifying the underlying condition guarantees a third occurrence — the condition has not changed, only the lock has been released

Stop Immediately — What Not to Do When Stuck

The instinct when a server is stuck is to keep hammering it with stop and start commands. This is the single most damaging thing you can do. Each retry call adds a new failed operation entry to the Azure Activity Log, making the timeline harder for the support engineer to read. In some control-plane implementations, a new failed operation can reset an internal timeout counter, meaning the stuck lock takes longer to clear. Do these things once, capture the output, and then stop.

⛔ Stop — Do Not Do Any of These

Do not repeatedly retry az postgres flexible-server stop or az postgres flexible-server start — run each command once to confirm the error, capture the output, and stop.

Do not attempt to delete and recreate the server — deletion operations are also blocked when the control-plane lock is held, and attempting to delete a stuck server can result in a server that is permanently in a deleting state with data inaccessible.

Do not attempt to scale up the SKU during the stuck state — scale operations are management-plane operations that will also be rejected and can add additional failed operations to the Activity Log.

Do not wait hours before opening a support ticket — the longer you wait, the shorter the evidence window in the Activity Log and the more context is lost from the backend monitoring systems.

Triage in 10 Minutes — Confirm the Lock and Collect Evidence

Run these commands once each. The goal is to confirm the stuck state, determine whether a platform incident is involved, and collect the correlation ID that the Microsoft support team needs to locate the specific stuck internal workflow.

1

azure.status.microsoft.com and portal → Service Health

Check for an active regional platform incident — 2 minutes

Open azure.status.microsoft.com and look for active incidents affecting Azure Database for PostgreSQL in your region. Also open Service Health in the Azure portal — this shows incidents specific to your subscription. A regional capacity constraint or zonal outage (documented cases include burstable SKU capacity issues in East US and other regions) can cause mass stuck-state events that resolve without any customer action once the incident closes. If you see an active incident, note the incident reference and include it in your support ticket.

2

Get current server state and collect the Correlation ID

Run the commands below. The most critical output is the Correlation ID from the Activity Log — a GUID that allows the backend team to find the specific stuck workflow instantly. Copy it exactly and save it before opening the ticket.

Azure CLI — Triage commands (run each ONCE, capture all output)SERVER="psql-xxx" # replace with your server name
RG="your-resource-group"

# 1. Get current server state az postgres flexible-server show \
  --name "${SERVER}" --resource-group "${RG}" \
  --query "{State:state,SKU:sku.name,Version:version,HA:highAvailability.mode}" \
  --output json

# 2. Get resource ID for Activity Log query RID=$(az postgres flexible-server show \
  --name "${SERVER}" --resource-group "${RG}" \
  --query id -o tsv)

# 3. Get Activity Log — COPY THE correlationId FROM THE OUTPUT az monitor activity-log list \
  --resource-id "${RID}" --max-events 20 \
  --query "[].{Time:eventTimestamp,Op:operationName.value,Status:status.value,CorrelationId:correlationId}" \
  --output table

# 4. Attempt stop ONCE — capture the full error, do NOT retry az postgres flexible-server stop \
  --name "${SERVER}" --resource-group "${RG}" 2>&1
# Expected error: (SeverBusyWithOtherOperation) Cannot perform 'Stop' server operation...

# 5. Check resource locks (rule out customer-created locks as a factor) az resource lock list \
  --resource-name "${SERVER}" \
  --resource-group "${RG}" \
  --resource-type "Microsoft.DBforPostgreSQL/flexibleServers" \
  --output table

3

PostgreSQL server → Help → Resource Health

Check Resource Health for PlatformInitiated events

Navigate to your PostgreSQL server in the portal → Help → Resource Health. This shows a timeline of availability events including PlatformInitiated downtime entries. Look for any events around the time the server got stuck. Screenshot any PlatformInitiated events — these are additional evidence of a control-plane issue for your support ticket and are especially useful for demonstrating the recurring pattern if a similar event appears on both the first and second occurrence dates.

4

PostgreSQL server → Help → Diagnose and solve problems

Run built-in diagnostics — saves support back-and-forth

Navigate to your server → Help → Diagnose and solve problems. Search for and run Availability and Connectivity and Long-Running Operations if available. Screenshot the results. These diagnostics sometimes surface hidden control-plane operations not visible in the Activity Log, and including them in your support ticket reduces the number of follow-up information requests from the support engineer.

Open a Severity A Support Ticket — Recurring Incident Template

Because this is the second occurrence in a month, frame the ticket explicitly as a recurring incident investigation, not just an unblock request. A ticket framed only as "please fix the stuck server" will result in the support engineer clearing the lock and closing the ticket — which is exactly what happened after the first occurrence, and it did not prevent the second. A ticket framed as a recurring incident investigation triggers a different internal workflow where the engineer is expected to identify and document the root cause, not just restore service.

ℹ Open here: aka.ms/azuresupport → Technical → Azure Database for PostgreSQL → Severity A

If you do not have a paid support plan, select Basic when prompted — Basic plan still covers service outages and stuck provisioning operations. Even if the portal shows only "Billing" or "Technical" as options, select Technical and describe the server as unavailable due to a platform-side lock. Do not attempt to open the ticket as a billing or advisory question — it will be routed to the wrong team and take significantly longer.

Support Ticket — Recurring Incident Template (copy, fill in, submit as Severity A)SUBJECT: [RECURRING — 2nd time in 30 days] Azure PostgreSQL Flexible Server stuck
in Starting — SeverBusyWithOtherOperation blocks all management operations

SEVERITY: A — Production database unavailable, second occurrence in 1 month

== SERVER DETAILS ==
Server name: psql-xxx
Resource group: [your-rg]
Subscription ID: [your-subscription-id]
Region: [e.g. West Europe]
SKU / tier: [e.g. Standard_B2ms Burstable / Standard_D4ds_v5 General Purpose]
PostgreSQL version: [e.g. 16]
High Availability: Enabled / Disabled
Read replicas: Yes / No

== CURRENT STATE ==
Provisioning state: Starting (stuck)
Stuck since: [timestamp UTC]
Impact: Database connections all failing, applications unavailable

== THIS IS A RECURRING INCIDENT ==
First occurrence date: [date ~1 month ago]
First occurrence resolution: Resolved by [support/platform team] — no root cause shared
Second occurrence date: [today's date]
Time between incidents: approximately 30 days
Pattern: same error, same server, same SeverBusyWithOtherOperation behaviour

== KEY EVIDENCE (CORRELATION IDs) ==
Activity Log — Correlation ID of stuck operation: [PASTE GUID]
Activity Log — Operation name: [e.g. "Start server"]
Activity Log — Timestamp: [e.g. 2026-07-08T08:19:00Z]
Resource Health — PlatformInitiated event: [Yes/No — paste details if yes]
Activity Log — No InProgress events visible: [Yes/No]

== EXACT CLI ERROR ==
(SeverBusyWithOtherOperation) Cannot perform 'Stop' server operation because
server 'psql-xxx' is busy processing other operation.
Code: SeverBusyWithOtherOperation
Message: Cannot perform 'Stop' server operation because server 'psql-xxx'
is busy processing other operation.

== WHAT WE TRIED ==
- az postgres flexible-server stop: SeverBusyWithOtherOperation
- Portal Stop button: greyed out
- Resource locks: none (az resource lock list returns empty)
- Did NOT retry to avoid adding noise to Activity Log

== REQUESTS (BOTH REQUIRED FOR RECURRING CASE) ==
1. Immediate: Clear the stuck backend operation and restore server to stable state
2. Investigation: Identify the root cause of this recurring pattern — why has
the same server got stuck twice in 30 days?
3. Documentation: Share the root cause finding with us so we can take
preventive action on our side
4. Recommendation: Advise on server configuration, SKU, or maintenance window
changes to prevent a third occurrence

Figure 2 — Support escalation and recovery flow for a recurring stuck state

Open the ticket immediately — triage should take no more than 10 minutes. Run PITR in parallel if production cannot wait for support to respond.

Emergency Recovery via PITR — If You Cannot Wait for Support

If your database is production-critical and waiting for Microsoft support to respond (typically 1–4 hours for Severity A) is not acceptable, you can initiate a Point-in-Time Restore (PITR) to create a new server from a recent backup. This gives you a working database quickly while support resolves the original stuck server in the background. Note that the restored server has a different hostname — connection strings must be updated across all applications.

If the PITR restore attempt itself fails with LocationRestricted or management errors, this means the control-plane lock is also blocking restore operations. Include this failure in your support ticket — it signals that the lock is more pervasive than a simple stuck workflow and requires a higher-priority internal escalation.

Azure CLI — PITR restore to a new server as emergency recovery# Find the available restore window az postgres flexible-server show \
  --name "${SERVER}" --resource-group "${RG}" \
  --query "{EarliestRestore:earliestRestoreDate,SKU:sku.name}" \
  --output json

# Restore to a new server — use a timestamp you know was before the stuck state # Adjust --restore-time to a safe point (e.g. last known good state) az postgres flexible-server restore \
  --name "${SERVER}-recovery" \
  --resource-group "${RG}" \
  --source-server "${SERVER}" \
  --restore-time "2026-07-08T07:00:00Z" \
  --sku-name "Standard_D4ds_v5" \
  --tier "GeneralPurpose"

# If restore fails with LocationRestricted or SeverBusyWithOtherOperation: # Add this note to your support ticket — lock is deeper than a simple stuck workflow # Request backend-initiated restore and document the PITR failure as additional evidence

The Six Root Causes Behind Recurring Stuck States

These are the confirmed root causes from documented Microsoft Q&A resolutions, support case histories, and the Azure PostgreSQL engineering blog. A recurring incident almost always involves one of these. Ask your support engineer explicitly which of these was identified for your server — and if they cannot confirm, request backend log evidence before closing the ticket.

Root Cause	How to Detect	Fix to Request / Apply
Platform maintenance window collision Azure background maintenance runs concurrently with a customer-triggered start/stop	Resource Health shows PlatformInitiated events at the same timestamp as the stuck state. Both incidents happened at similar times of day or week.	Configure a dedicated maintenance window in the portal at a time that does not overlap with any automation or peak usage windows. Request support confirm whether platform maintenance triggered both incidents.
Storage full — WAL accumulation from inactive replication slots PostgreSQL retains WAL for inactive logical replication slots until disk is full	Storage metrics at or near 100% before the stuck event. Log entries referencing "no space left on device". Support backend confirmed WAL accumulation in prior resolved cases.	Enable Auto Storage Growth in Compute + Storage settings. Drop or monitor inactive replication slots. Enable storage alert at 75%.
CPU credit exhaustion (burstable B-series SKU) B-series VMs accumulate CPU credits when idle; startup requires credit balance	Server SKU is Standard_B*ms. CPU credits metric was at zero or near-zero when the start was attempted. Previous incident also on a B-series SKU.	Upgrade to a General Purpose (D-series) SKU — dedicated vCores, no credit model, more reliable for start sequences. Do not use burstable SKUs for databases that need reliable starts.
7-day auto-start collision with automation Azure auto-starts a stopped server after 7 days, colliding with customer automation	Incidents separated by approximately 7-day multiples. Server is stopped regularly and incidents occur on the expected auto-start day.	Ensure automation start commands include a pre-check for server state. Schedule automation slightly after the expected auto-start window. Consider upgrading to a non-stopped operational model.
Zonal outage or regional capacity constraint Azure regional infrastructure issue affecting specific SKU families or AZs	Active Azure Status incident in your region. Multiple customers reporting the same issue simultaneously. SKU type matches the affected family (e.g. burstable in East US).	Nothing customer-side can accelerate recovery. Monitor Azure Status. Consider Availability Zone redundant deployment if the region supports it, to survive single-AZ failures.
Corrupted internal metadata or expired SAS key Backend storage access credentials or configuration files fail silently during start	No obvious customer-side cause. Support backend investigation found a reserved space file or expired SAS key. Recurring at irregular intervals.	This requires backend mitigation only — the support team must reset the SAS key or remove the corrupt file. Request that support confirm whether this was the root cause and verify it has been resolved for both occurrences.

Post-Incident Hardening Checklist — Stop the Third Occurrence

After the server is restored to a stable state, do not simply restart operations. Work through this checklist before declaring the incident closed. Every item addresses a confirmed recurrence trigger.

Upgrade from burstable (B-series) to General Purpose (D-series) SKU if the server is on Standard_B*ms. The burstable credit model is not designed for databases that need reliable, consistent startup sequences. Even a Standard_D2ds_v4 is significantly more reliable than B2ms for this workload pattern.

Configure a dedicated maintenance window in the portal (PostgreSQL server → Maintenance). Set it to a window at least 4 hours away from any automation or peak usage. This prevents platform maintenance from running concurrently with your operations.

Enable Auto Storage Growth (Compute + Storage → Auto scale storage). This prevents the server from getting stuck at startup due to a full disk. Set an alert at 75% storage utilisation as a secondary guard.

Check for inactive logical replication slots. Connect to the database once it is back online and run the query below. Any slot with no active consumer is accumulating WAL. Drop slots that are not actively being consumed.

Request written root cause confirmation from support before closing the ticket. "The server has been fixed" is not sufficient closure for a recurring incident. Ask for the specific internal root cause that was identified for both occurrences.

Add server state monitoring (see Section 8) so that the next time the server enters Starting or Stopping for more than 15 minutes, you receive an immediate alert rather than discovering it after applications have been failing for an extended period.

Evaluate whether High Availability should be enabled. Zone-redundant HA provides a standby replica in a different availability zone. While it does not prevent a stuck state, it can provide faster recovery in some scenarios because the standby can be promoted independently of the primary's stuck state.

psql — Check for inactive replication slots after server is back online-- Connect to your database as the admin user
-- Run this FIRST after the server recovers

SELECT
  slot_name,
  plugin,
  active,
  pg_size_pretty(pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn)) AS retained_wal
FROM pg_replication_slots
ORDER BY retained_wal DESC;

-- If any slot shows active = false AND retained_wal is large: -- DROP REPLICATION SLOT 'slot_name_here'; -- This reclaims the retained WAL space immediately

-- Also check current disk usage SELECT
  datname,
  pg_size_pretty(pg_database_size(datname)) AS db_size
FROM pg_database
ORDER BY pg_database_size(datname) DESC;

Monitoring Setup — Alert Before the Next Stuck State Happens

The right time to find out your PostgreSQL server is stuck is not when applications start failing — it is within minutes of the server entering a transitional state. These monitoring configurations alert before users are impacted.

Azure CLI — Set up alerts for stuck state and high storage utilisation# Replace variables SERVER="psql-xxx"
RG="your-resource-group"
EMAIL="ops-team@yourdomain.com"
SUBSCRIPTION="your-subscription-id"

# Get the server resource ID RID=$(az postgres flexible-server show --name "${SERVER}" --resource-group "${RG}" --query id -o tsv)

# Create an Action Group for email alerting az monitor action-group create \
  --resource-group "${RG}" \
  --name "pg-alert-group" \
  --short-name "pgalert" \
  --action email ops-email "${EMAIL}"

AG_ID=$(az monitor action-group show --resource-group "${RG}" --name "pg-alert-group" --query id -o tsv)

# Alert 1: Storage utilisation above 75% (early warning before disk full) az monitor metrics alert create \
  --name "pg-storage-high" \
  --resource-group "${RG}" \
  --scopes "${RID}" \
  --condition "avg storage_percent > 75" \
  --window-size 5m \
  --evaluation-frequency 5m \
  --severity 2 \
  --action "${AG_ID}" \
  --description "PostgreSQL storage above 75% — risk of stuck start on next restart"

# Alert 2: Activity Log alert when server Start/Stop operation begins # This fires when the operation is accepted — gives you early visibility az monitor activity-log alert create \
  --name "pg-start-stop-activity" \
  --resource-group "${RG}" \
  --scope "${RID}" \
  --condition "category=Administrative and operationName=Microsoft.DBforPostgreSQL/flexibleServers/start/action" \
  --action-group "${AG_ID}" \
  --description "PostgreSQL start operation triggered — monitor for stuck state"

# Alert 3: CPU credits remaining (for burstable SKUs only) # If credits drop to zero, next start may hang az monitor metrics alert create \
  --name "pg-cpu-credits-low" \
  --resource-group "${RG}" \
  --scopes "${RID}" \
  --condition "avg cpu_credits_remaining < 10" \
  --window-size 5m \
  --evaluation-frequency 5m \
  --severity 2 \
  --action "${AG_ID}" \
  --description "CPU credits critically low — consider upgrading to General Purpose SKU"

Key Takeaways

✓A second occurrence in 30 days means an unresolved underlying condition — not bad luck. The support ticket for the second occurrence must explicitly request root cause investigation, not just an unblock.

✓SeverBusyWithOtherOperation on Stop and Start means a backend control-plane lock. You cannot clear it from the portal or CLI. Do not retry commands — run each once, capture the output, and open a Severity A support ticket immediately.

✓The Correlation ID from the Activity Log is the single most important item for the support engineer — it lets them locate the specific stuck internal workflow without manual searching. Copy it before opening the ticket.

✓Frame the ticket explicitly as a recurring incident. State the first occurrence date and resolution. Request root cause documentation as part of the closure criteria, not just service restoration.

✓If production cannot wait for support (1–4 hours for Severity A), initiate a PITR restore to a new server in parallel. If PITR also fails, include this in the ticket as additional evidence of a deeper management-plane lock.

✓The 7-day auto-start is a documented Azure behaviour — a stopped Flexible Server is automatically restarted after 7 days for maintenance. If your stop/start pattern creates a weekly cadence, this auto-start can collide with your automation. Check whether both incidents align with 7-day multiples.

✓After recovery, upgrade from burstable (B-series) to General Purpose (D-series) if applicable, enable Auto Storage Growth, configure a dedicated maintenance window, check for inactive replication slots, and enable storage and activity alerts before the next potential incident.

Frequently Asked Questions

Why did the first support fix not prevent the second occurrence?

Because clearing the stuck backend lock does not address the condition that caused the lock to form. If the server got stuck because of a maintenance window collision, burstable credit exhaustion, or storage filling up — all of those conditions still exist after the lock is cleared. The server starts working again, the condition rebuilds over the following weeks, and the next triggering event causes the same stuck state. The only way to prevent recurrence is to identify and fix the underlying condition during the same support interaction, not as a separate follow-up.

The support engineer just cleared the lock and closed the ticket without sharing a root cause — what should I do?

Reopen the ticket or open a new one explicitly requesting root cause. Write: "This is the second occurrence in 30 days. Please confirm the specific internal root cause that was identified — not just the mitigation applied. We need to understand what recurring condition is making this server susceptible to this failure before we can take preventive action." If the engineer cannot identify the root cause, ask them to escalate to the Azure Database for PostgreSQL engineering team for a deeper investigation. A ticket framed this way is much harder to close without a substantive response.

Is there a way to check if the 7-day auto-start is involved?

Yes. Look at the Resource Health blade for your server and examine the timeline of PlatformInitiated events. The 7-day auto-start generates a PlatformInitiated restart entry. If you see this entry at or near the time of the stuck state, and the two incidents are separated by approximately 7 days or multiples thereof, the auto-start is a contributing trigger. You can also check the Activity Log for a "Start server" operation initiated by "Microsoft Azure" (platform) rather than your own user account or service principal — this distinguishes a platform-initiated start from a customer-triggered one.

Would switching from Burstable to General Purpose definitely fix this?

It addresses one confirmed recurrence trigger — CPU credit exhaustion. If that is the root cause, yes. If the root cause is a maintenance window collision, storage issue, or platform infrastructure problem, upgrading the SKU alone will not prevent recurrence. The SKU upgrade is one item on the hardening checklist, not the only item. Run all of the post-incident checks in Section 7 in parallel with requesting a root cause from support.

Are there storage charges while the server is stuck in Starting?

Yes. Storage charges (provisioned storage and backup storage) continue regardless of the server's operational state. Compute charges are paused only when the server is in a cleanly Stopped state. A server stuck in Starting is technically not in a fully stopped state — whether compute is billed during a stuck transitional state depends on the specific scenario and the platform-side view of the server. If the stuck state persists for hours, include the duration in your support ticket. Microsoft support can review billing concerns but typically cannot waive charges retroactively — the priority is restoring the service, and billing questions should follow as a separate case if needed.

FAVORITE Azure Engineering, AI & Real-World Fixes

Search This Blog