Skip to content
|
Kong Gateway
3.6.x
github-edit-pageEdit this page
report-issueReport an issue
You are browsing documentation for an older version. See the latest documentation here.

Resource Sizing Guidelines

This document discusses the performance characteristics of Kong Gateway, and offers recommendations on sizing for resource allocation based on expected Kong Gateway configuration and traffic patterns.

These recommendations are a baseline guide only.

Specific tuning or benchmarking efforts should be undertaken for performance-critical environments.

General resource guidelines

Kong Gateway resources

Kong Gateway is designed to operate in a variety of deployment environments. It has no minimum system requirements to operate.

Resource requirements vary substantially based on configuration. The following high-level matrices offer a guideline for determining system requirements based on overall configuration and performance requirements.

Consider the following simplified examples, where latency and throughput requirements are considered on a per-node basis. This table has rough usage requirement estimates:

Size Number of Configured Entities Latency Requirements Throughput Requirements Usage Pattern
Development < 100 < 100 ms < 500 RPS Dev/test environments; latency-insensitive gateways
Small < 1000 < 20 ms < 2500 RPS Production clusters; greenfield traffic deployments
Medium < 10000 < 10 ms < 10000 RPS Mission-critical clusters; legacy & greenfield traffic; central enterprise-grade gateways
Large < 50000+ < 10 ms < 10000 RPS Mission-critical clusters; legacy & greenfield traffic; central enterprise-grade gateways

Database resources

We do not provide any hard numbers for database sizing (DB sizing), as it depends on your particular setup. Sizing varies based on:

  • Traffic
  • Number of nodes
  • Enabled features: for example, if rate limiting uses a database or Redis
  • Number and rate of change of configured entities
  • The rate at which Kong Gateway processes are started and restarted within the cluster
  • The size of Kong Gateway’s in-memory cache

Kong Gateway intentionally relies on the database as little as possible. To access configuration, Kong Gateway executes a spiky access pattern to its backing database. This means that Kong Gateway only reads configuration from the database when a node first starts, or configuration for a given entity changes.

Everything in the database is meant to be read infrequently and held in memory as long as possible. Therefore, database resource requirements are lower than those of compute environments running Kong Gateway.

Query patterns are typically simple and follow schema indexes. Provision sufficient database resources in order to handle spiky query patterns.

There are settings that you can adjust to keep database access minimal (also see in-memory caching), or keep Kong Gateway operational if the DB is down for maintenance. If you choose to keep the database operational during downtime, vitals data is not written to the database during this time.

Cluster resource allocations

Based on the expected size and demand of the cluster, we recommend the following resource allocations as a starting point:

Size CPU RAM Typical Cloud Instance Sizes
Development 1-2 cores 2-4 GB AWS: t3.medium
GCP: n1-standard-1
Azure: Standard A1 v2
Small 1-2 cores 2-4 GB AWS: t3.medium
GCP: n1-standard-1
Azure: Standard A1 v2
Medium 2-4 cores 4-8 GB AWS: m5.large
GCP: n1-standard-4
Azure: Standard A1 v4
Large 8-16 cores 16-32 GB AWS: c5.xlarge
GCP: n1-highcpu-16
Azure: F8s v2

We strongly discourage the use of throttled cloud instance types (such as the AWS t2 or t3 series of machines) in large clusters, as CPU throttling would be detrimental to Kong Gateway’s performance. We also recommend testing and verifying the bandwidth availability for a given instance class. Bandwidth requirements for Kong Gateway depend on the shape and volume of traffic flowing through the cluster.

In-memory caching

We recommend defining the mem_cache_size configuration as large as possible, while still providing adequate resources to the operating system and any other processes running adjacent to Kong Gateway. This configuration allows Kong Gateway to take maximum advantage of the in-memory cache, and reduce the number of trips to the database.

Each Kong Gateway worker process maintains its own memory allocations, and must be accounted for when provisioning memory. By default, one worker process runs per number of available CPU cores. We recommend allowing for around 500MB of memory allocated per worker process.

For example, on a machine with 4 CPU cores and 8 GB of RAM available, we recommend allocating between 4-6 GB to cache via the mem_cache_size directive, depending on what other processes are running alongside Kong Gateway.

Plugin queues

Several plugins that are distributed with Kong Gateway use internal, in-memory queues to decouple production of data from the transmission to an upstream server. These queues reduce the number of concurrent requests that are made to an upstream server under high load conditions and provide for buffering during temporary network and upstream outages. For more information about Kong Gateway’s internal queueing system, see About Plugin Queuing.

As queues use main memory to store queued entries, it is important to understand how many queues exist in the system and how many entries they can hold in terms of their capacity configuration.

Most plugins use one queue per plugin instance, with the exception of the HTTP Log plugin, which uses one queue per log server upstream configuration.

The queue.max_entries configuration parameter determines how many entries can be waiting for transmission in a given queue. Once this limit is reached, the oldest entry is removed when a new entry is enqueued. While it is not possible to precisely predict how much memory a single queue entry will occupy for a given plugin and in a particular configuration, estimates can be made based on the amount of data that is actually transmitted to the upstream server.

In larger configurations, it is advisable to experimentally determine the memory requirements of queues by running Kong Gateway in a test environment and observing its memory consumption while plugin upstream servers are unavailable, forcing queues to reach the configured limits.

The default value of 10,000 entries for the queue.max_entries should provide for enough buffering in many installations while keeping the maximum memory usage of queues at reasonable levels.

Scaling dimensions

Kong Gateway is designed to handle large volumes of request traffic and proxying requests with minimal latency. Understanding how various configuration scenarios impacts request traffic, and the Kong Gateway cluster itself, is a crucial step in successfully deploying Kong Gateway.

Kong Gateway measures performance in the following dimensions:

  • Latency refers to the delay between the downstream client sending a request and receiving a response. Kong Gateway measures latency introduced into the request in terms of microseconds or milliseconds. Increasing the number of Routes and Plugins in a Kong Gateway cluster increases the amount of latency that’s added to each request.
  • Throughput refers to the number of requests that Kong Gateway can process in a given time span, typically measured in seconds or minutes.

These dimensions have an inversely proportional relationship when all other factors remain the same: decreasing the latency introduced into each request allows the maximum throughput in Kong Gateway to increase, as there is less CPU time spent handling each request, and more CPU available for processing traffic as a whole. Kong Gateway is designed to scale horizontally to be able to add more overall compute power for configurations that add substantial latency into requests, while needing to meet specific throughput requirements.

Kong Gateway’s maximum throughput is a CPU-bound dimension, and minimum latency is memory-bound.

  • Latency-sensitive workload: making more memory available for database caching is more beneficial than adding more compute power to the cluster.
  • Throughput-sensitive workload: these workloads are dependant on both adequate memory and CPU resources, but adding more compute power by scaling Kong Gateway vertically or horizontally is the better choice, as it provides near-unlimited throughput capacity. In this scenario, adding more cache memory would not increase maximum throughput by much.

When Kong Gateway is operating in hybrid mode with a large number of configuration entities (routes, services, etc.), it can benefit from the dedicated configuration processing option. When enabled, certain CPU-intensive steps of the data plane reconfiguration operation are offloaded to a dedicated worker process. This reduces proxy latency during reconfigurations at the cost of a slight increase in memory usage. The benefits of this mechanism are most apparent with configurations consisting of more than a thousand configuration entities. See the configuration reference for dedicated_config_processing for more information.

Performance benchmarking and optimization as a whole is a complex exercise that must account for a variety of factors, including those external to Kong Gateway, such as the behavior of upstream services, or the health of the underlying hardware on which Kong Gateway is running.

Performance characteristics

There are a number of factors that impact Kong Gateway’s performance, including:

  • Number of configured Routes and Services: Increasing the count of Routes and Services on the cluster requires more CPU to evaluate the request. However, Kong Gateway’s request router can handle running at large scale. We’ve seen clusters of Kong Gateway nodes serving tens of thousands of Routes with minimal impact to latency as a result of request route evaluation.

  • Number of configured Consumers and Credentials: Consumer and credential data is stored in Kong Gateway’s datastore. Kong Gateway caches this data in memory to reduce database load and latency during request processing. Increasing the count of Consumers and Credentials requires more memory available for Kong Gateway to hold data in cache. If there is not enough memory available to cache all requested database entities, request latency increases as Kong Gateway needs to query the database more frequently to satisfy requests.

  • Number of configured Plugins: Increasing the count of Plugins on the cluster requires more CPU to iterate through plugins during request processing. Executing plugins comes with a varying cost depending on the nature of the plugin. For example, a lightweight authentication plugin like key-auth requires less resource availability than a plugin that performs complex transformations of an HTTP request or response.

  • Cardinality of configured Plugins: Cardinality is the number of distinct plugin types that are configured on the cluster. For example, a cluster with one each of ip-restriction, key-auth, bot-detection, rate-limiting, and http-log plugins has a higher plugin cardinality than a cluster with one thousand rate-limiting plugins applied at the route level. With each additional plugin type added to the cluster, Kong Gateway spends more time evaluating whether to execute a given plugin for a given request. Increasing the cardinality of configured plugins requires more CPU power, as the process to evaluate plugins is a CPU-constrained task.

  • Request and response size: Requests with large HTTP bodies, either in the request or response, take longer to process, as Kong Gateway must buffer the request to disk before proxying it. This allows Kong Gateway to handle a large volume of traffic without running out of memory, but the nature of buffered requests can result in increased latency.

  • Number of configured Workspaces: Increasing the count of Workspaces on the cluster requires more CPU to evaluate each request, and more memory available for the cache to hold Workspace configuration and metadata. The impact of increasing the number of Workspaces on the cluster is also affected by the cardinality of configured plugins on the cluster. There is an exponential impact on request throughput capacity within the cluster as the cardinality of plugins and the number of Workspaces increases.

More information