Slow Consumers

In core NATS, consumers that cannot keep up are handled differently from many other messaging systems: NATS favors the approach of protecting the system as a whole over accommodating a particular consumer to ensure message delivery.

What is a slow consumer?

A slow consumer is a subscriber that cannot keep up with the message flow delivered from the NATS server. This is a common case in distributed systems because it is often easier to generate data than it is to process it. When consumers cannot process data fast enough, back pressure is applied to the rest of the system. NATS has mechanisms to reduce this back pressure.

NATS identifies slow consumers in the client or the server, providing notification through registered callbacks, log messages, and statistics in the server’s monitoring endpoints.

What happens to slow consumers?

When detected at the client, the application is notified and messages are dropped to allow the consumer to continue and reduce potential back pressure. When detected in the server, the server will disconnect the connection with the slow consumer to protect itself and the integrity of the messaging system.

A client can detect it is a slow consumer on a local connection and notify the application through use of the asynchronous error callback. It is better to catch a slow consumer locally in the client rather than to allow the server to detect this condition. This example demonstrates how to define and register an asynchronous error handler that will handle slow consumer errors.

With this example code and default settings, a slow consumer error would generate output something like this:

Note that if you are using a synchronous subscriber, will also return an error indicating there was a slow consumer and messages have been dropped.

When the server initiates a slow consumer error, you’ll see the following in the server output:

The server will also keep count of the number of slow consumer errors encountered, available through the monitoring varz endpoint in the slow_consumers field.

Apart from using or optimizing your consuming application, there are a few options available: scale, meter, or tune NATS to your environment.

Scaling with queue subscribers

This is ideal if you do not rely on message order. Ensure your NATS subscription belongs to a queue group, then scale as required by creating more instances of your service or application. This is a great approach for microservices - each instance of your microservice will receive a portion of the messages to process, and simply add more instances of your service to scale. No code changes, configuration changes, or downtime whatsoever.

Create a subject namespace that can scale

You can distribute work further through the subject namespace, with some forethought in design. This approach is useful if you need to preserve message order. The general idea is to publish to a deep subject namespace, and consume with wildcard subscriptions while giving yourself room to expand and distribute work in the future.

For a simple example, if you have a service that receives telemetry data from IoT devices located throughout a city, you can publish to a subject namespace like , Sensors.South, Sensors.East and . Initially, you’ll subscribe to Sensors.> to process everything in one consumer. As your enterprise grows and data rates exceed what one consumer can handle, you can replace your single consumer with four consuming applications to subscribe to each subject representing a smaller segment of your data. Note that your publishing applications remain untouched.

Meter the publisher

Tune NATS through configuration

The NATS server can be tuned to determine how much data can be buffered before a consumer is considered slow, and some officially supported clients allow buffer sizes to be adjusted. Decreasing buffer sizes will let you identify slow consumers more quickly. Increasing buffer sizes is not typically recommended unless you are handling temporary bursts of data. Often, increasing buffer capacity will only postpone slow consumer problems.

The NATS server has a write deadline it uses to write to a connection. When this write deadline is exceeded, a client is considered to have a slow consumer. If you are encountering slow consumer errors in the server, you can increase the write deadline to buffer more data.

The write_deadline configuration option in the NATS server configuration file will tune this:

Tuning this parameter is ideal when you have bursts of data to accommodate. Be sure you are not just postponing a slow consumer error.

Client Configuration

Most officially supported clients have an internal buffer of pending messages and will notify your application through an asynchronous error callback if a local subscription is not catching up. Receiving an error locally does not necessarily mean that the server will have identified a subscription as a slow consumer.

This buffer can be configured through setting the pending limits after a subscription has been created:

The default subscriber pending message limit is , and the default subscriber pending byte limit is 65536*1024

If the client reaches this internal limit, it will drop messages and continue to process new messages. This is aligned with NATS at most once delivery. It is up to your application to detect the missing messages and recover from this condition.