Skip to main content
cancel
Showing results for 
Search instead for 
Did you mean: 

Data Days is here! Join us now for 60+ days of learning, challenges, and connection. Learn more

Reply
carlton7372
Helper III
Helper III

Azure Fabric Spark Sessions Taking Very Long Time to Spin Up

Hi Community,

 

I have researched that it's possible to configure Spark Sessions to spin up quickly. At the moment, it's taking up to 8mins for a Standard Session to spin up.

Can someone let me know what is needed to make Sparks not so long to spin up?

 

carlton7372_0-1760286850199.png

 

1 ACCEPTED SOLUTION
AntoineW
Super User
Super User

Hello @carlton7372,

 

In Microsoft Fabric, every time you run a Notebook, Dataflow Gen2, or any Spark-based job, the service provisions an isolated Spark session on-demand inside your capacity (F SKU).
The “spin-up time” — typically 3–8 minutes — is the time Fabric needs to:

  1. Allocate Spark compute resources,

  2. Initialize cluster dependencies and libraries,

  3. Mount OneLake storage and environment variables,

  4. Start your session driver and executors.

If you’re seeing 8+ minutes regularly, it usually indicates that Spark pools are not being reused efficiently or capacity resources are constrained.

 

Here are the most common causes:

1. Cold Start (No Warm Pool)

Fabric currently provisions Spark sessions on demand. If there’s no active Spark session in your capacity, the first job triggers a cold start — allocating new containers and initializing runtime images.

 

2. Capacity Idle Timeout

Each Fabric capacity (F SKU) auto-scales down when idle to save resources.
After 20–30 minutes of inactivity, Spark runtimes are deallocated, meaning your next run must spin up new containers again.

 

3. Large Session Configuration

Launching a Standard or Large session type allocates more executors and memory, increasing startup latency.
Unless you’re running heavy transformations or ML jobs, a Small session usually suffices and spins up faster.

 

4. Competing Workloads on the Same Capacity

If multiple users or Dataflows are consuming your F SKU at the same time, Fabric’s resource scheduler may queue new sessions until enough CUs are available.

 

5. Custom Libraries or Initialization Overhead

If your Notebook installs additional packages using %pip install or mounts external data sources, that also adds delay to the effective “ready-to-run” time.

 

Source : 

https://learn.microsoft.com/en-us/fabric/data-engineering/spark-compute

 

Hope it can help you ! 

Best regards,

Antoine

View solution in original post

1 REPLY 1
AntoineW
Super User
Super User

Hello @carlton7372,

 

In Microsoft Fabric, every time you run a Notebook, Dataflow Gen2, or any Spark-based job, the service provisions an isolated Spark session on-demand inside your capacity (F SKU).
The “spin-up time” — typically 3–8 minutes — is the time Fabric needs to:

  1. Allocate Spark compute resources,

  2. Initialize cluster dependencies and libraries,

  3. Mount OneLake storage and environment variables,

  4. Start your session driver and executors.

If you’re seeing 8+ minutes regularly, it usually indicates that Spark pools are not being reused efficiently or capacity resources are constrained.

 

Here are the most common causes:

1. Cold Start (No Warm Pool)

Fabric currently provisions Spark sessions on demand. If there’s no active Spark session in your capacity, the first job triggers a cold start — allocating new containers and initializing runtime images.

 

2. Capacity Idle Timeout

Each Fabric capacity (F SKU) auto-scales down when idle to save resources.
After 20–30 minutes of inactivity, Spark runtimes are deallocated, meaning your next run must spin up new containers again.

 

3. Large Session Configuration

Launching a Standard or Large session type allocates more executors and memory, increasing startup latency.
Unless you’re running heavy transformations or ML jobs, a Small session usually suffices and spins up faster.

 

4. Competing Workloads on the Same Capacity

If multiple users or Dataflows are consuming your F SKU at the same time, Fabric’s resource scheduler may queue new sessions until enough CUs are available.

 

5. Custom Libraries or Initialization Overhead

If your Notebook installs additional packages using %pip install or mounts external data sources, that also adds delay to the effective “ready-to-run” time.

 

Source : 

https://learn.microsoft.com/en-us/fabric/data-engineering/spark-compute

 

Hope it can help you ! 

Best regards,

Antoine

Helpful resources

Announcements
Fabric Data Days is here Carousel

Fabric Data Days 2026

Don't miss out on Data Days, June 15 through August 7. Learn Fabric, Power BI, SQL, AI and more.

June Fabric Update Carousel

Fabric Monthly Update - June 2026

Check out the June 2026 Fabric update to learn about new features.