---
title: "How we raised Databricks cluster rate limits by 7.5×"
description: "Databricks’ Compute Platform team replaced fixed request-rate limits for cluster creation with adaptive concurrency limits to better handle long-running, variable-cost operations. Fixed limits turned away requests during healthy periods yet accepted too many when dependencies slowed, requiring manual tuning. Adaptive concurrency tracks active provisioning workflows with slots and adjusts the slot count based on congestion signals such as provisioning latency and thread pool saturation. The rollout analysis reported a 7.5× increase in default cluster-creation request-rate limits, more than 99.99% fewer rate-limit errors, and removal of over a hundred customer-specific overrides. The fixed API request-rate limit remains as a final safety guardrail against exceptionally high incoming traffic."
---

# How we raised Databricks cluster rate limits by 7.5×

[Databricks](https://yomu.fyi/company/databricks) · Heqing Ya, Tan Zhang, Ankit Batra, Abhishek Rai · Nov 18, 2026

**Type:** Problem & solution

## Summary

Databricks’ Compute Platform team replaced fixed request-rate limits for cluster creation with adaptive concurrency limits to better handle long-running, variable-cost operations. Fixed limits turned away requests during healthy periods yet accepted too many when dependencies slowed, requiring manual tuning. Adaptive concurrency tracks active provisioning workflows with slots and adjusts the slot count based on congestion signals such as provisioning latency and thread pool saturation. The rollout analysis reported a 7.5× increase in default cluster-creation request-rate limits, more than 99.99% fewer rate-limit errors, and removal of over a hundred customer-specific overrides. The fixed API request-rate limit remains as a final safety guardrail against exceptionally high incoming traffic.

## Context

Databricks’ Compute Platform team needed to protect long-running cluster-creation workflows from overload while maximizing capacity. Fixed request-rate limits were difficult to tune because cluster-creation operations vary in resource cost and duration, and dependency health changes unpredictably. Conservative limits turned away requests during healthy periods, while degradation could still cause retry storms that sustained overload even after traffic returned to normal.

## Approach / What changed

Adaptive concurrency limits that track active cluster-creation workflows with slots and adjust the slot count based on congestion signals such as provisioning latency and thread pool saturation.

## Takeaways

- Fixed request-rate limits are a poor model of capacity for long-running operations because one request can trigger variable amounts of background work and dependency health changes unpredictably.
- Adaptive concurrency limits track active operations with slots and adjust the slot count based on congestion signals, letting more work start when capacity is available and slowing the pace during degradation.
- Pairing concurrency limits with retry backoff and a limit on attempts helps reduce retry-storm pressure during recovery from a dependency slowdown.

## FAQ

### What is the difference between request-rate and concurrency limits?

A request-rate limit caps how many requests are accepted per unit time; a concurrency limit caps how many operations are active at once. Each active operation holds a slot until it finishes, including time spent waiting on dependencies.

### How do adaptive concurrency limits respond to system degradation?

When a cloud provider slows down, provisioning workflows hold their slots longer and accepted requests wait longer to launch. The controller also reduces the slot limit when congestion signals, such as provisioning latency or thread pool saturation, indicate pressure. As conditions recover, it cautiously allows more concurrent work.

### Why does Databricks still keep a fixed API request-rate limit?

The fixed API request-rate limit protects against exceptionally high incoming traffic and provides a request-rate boundary that can be communicated to customers. Adaptive concurrency limits work inside the service: capacity pressure queues accepted requests and delays provisioning. A successful API response does not guarantee an immediate cluster launch or completion within a fixed time.

**Tags:** [Concurrency](https://yomu.fyi/topic/concurrency), [Databricks](https://yomu.fyi/topic/databricks), [Rate Limiting](https://yomu.fyi/topic/rate-limiting), [Resilience](https://yomu.fyi/topic/resilience)

- Source: [Databricks](https://www.databricks.com/blog/how-we-raised-databricks-cluster-rate-limits-75)
- Source URL: https://www.databricks.com/blog/how-we-raised-databricks-cluster-rate-limits-75
- Ingested by Yomu: 2026-11-18T03:00:10.658Z

[Read original post](https://www.databricks.com/blog/how-we-raised-databricks-cluster-rate-limits-75)
