Flink Autotuning #

Flink Autotuning aims at fully automating the configuration of Apache Flink.

One of the biggest challenges with deploying new Flink pipelines is to write an adequate Flink configuration. The most important configuration values are:

  • memory configuration (heap memory, network memory, managed memory, JVM off-heap, etc.)
  • number of task slots

Memory Autotuning #

As a first step, we have tackled the memory configuration which, according to users, is the most frustrating part of the configuration process. The most important aspect of the memory configuration is the right-sizing of the various Flink memory pools. These memory pools include: heap memory, network memory, managed memory, and JVM off-heap memory settings. Non-optimal configuration of these pools can cause application crashes, or block large amounts of memory which remain unused.

How It Works #

With Flink Autoscaling and Flink Autotuning, all users need to do is set a max memory size for the TaskManagers, just like they would normally configure TaskManager memory. Flink Autotuning then automatically adjusts the various memory pools and brings down the total container memory size. It does that by observing the actual max memory usage on the TaskMangers or by calculating the exact number of network buffers required for the job topology. The adjustments are made together with Flink Autoscaling, so there is no extra downtime involved.

It is important to note that adjusting the container memory only works on Kubernetes and that the initially provided memory settings represent the maximum amount of memory Flink Autotuning will use. You may want to be more conservative than usual when initially assigning memory with Autotuning. We never go beyond the initial limits to ensure we can safely create TaskManagers without running into pod memory quotas or limits.

Getting Started #

Dry-run Mode #

As soon as Flink Autoscaling is enabled, Flink Autotuning will provide recommendations via events (e.g. Kubernetes events):

# Autoscaling needs to be enabled
job.autoscaler.enabled: true
# Disable automatic memory tuning (only get recommendations)
job.autoscaler.memory.tuning.enabled: false

Automatic Mode #

Automatic memory tuning via can be enabled by setting:

# Autoscaling needs to be enabled
job.autoscaler.enabled: true
# Turn on Autotuning and apply memory config changes
job.autoscaler.memory.tuning.enabled: true

Advanced Options #

Maximize Managed Memory #

Enabling the following option allows to return all saved memory as managed memory. This is beneficial when running with RocksDB to maximize its performance.

job.autoscaler.memory.tuning.maximize-managed-memory: true

Setting Memory Overhead #

Memory Autotuning uses a constant amount of memory overhead for heap and metaspace to allow the memory to grow beyond the determined maximum size. The default of 20% can be changed to 50% by setting:

job.autoscaler.memory.tuning.overhead: 0.5

Future Work #

Task Slots Autotuning #

The number of task slots are partially taken care by Flink Autoscaling which adjusts the task parallelism and hence changes the total number of slots and the number of TaskManagers.

In future versions of Flink Autotuning, we will try to further optimize the number of task slots depending on the number of tasks running inside a task slot.

JobManager Memory Tuning #

Currently, only TaskManager memory is adjusted.

RocksDB Memory Tuning #

Currently, if no managed memory is used, e.g. the heap-based state backend is used, managed memory will be set to zero by Flink Autotuning which helps save a lot of memory. However, if managed memory is used, e.g. via RocksDB, the configured managed memory will be kept constant because Flink currently lacks metrics to accurately measure the usage of managed memory.

Nevertheless, users already benefit from the resource savings and optimizations for heap, metaspace, and network memory. RocksDB users can solely focus their attention on configuring managed memory.

We already added an option to add all saved memory to the managed memory. This is beneficial when running with RocksDB to maximize the in-memory performance.