Skip to content
This repository has been archived by the owner on Nov 9, 2020. It is now read-only.

servicePauseResume

Tadaya Tsuyukubo edited this page Apr 21, 2017 · 3 revisions

Overview

A Xenon service instance consumes a small amount of memory to keep track of runtime context (operation queues, stats, subscribers, its URI entry in the host map). The total per service is about 300 bytes, so while tiny, for a fully indexed, highly available, scale out, asynchronous component, it can still potentially limit the total number of services instances (with each representing a configuration object or document in a control plane).

In order to make Xenon be disk-bound only, not memory bound, a on demand pause/resume exists, which periodically calculates service host memory use, and persists runtime service context on a dedicated "blob" index. The service runtime context is removed completely.

Requirements

The ServiceOption.ON_DEMAND_LOAD(ODL) must be enabled.

Usage guidelines

  • Frequently used, low latency services should NOT be ODL. They should be in memory, indexed or not indexed. However, they can only exist as many as memory allows.
  • Most document backend services and ALL tasks should be ODL

Note: Pause vs Stop
When ODL services have no soft state(subscribers or stats), simply they will be stopped(removed). For ODL services that have soft state, they will be paused(indexed), so their soft state will not be lost.

Pause

ServiceHost runs periodic maintenance and uses the setServiceRelativeMemoryLimit(ServiceHost.ROOT_PATH) API to determine the upper bound on its own memory use. The default bound is 50% of total JVM memory. During maintenance the host does the following

  • Computes the total size, in MB of memory utilized by caching, operation queues, and per service runtime context. Some defaults are used to estimate memory cost, based on numbers derived from memory analysis tools
  • Retrieves the LOW_WATERMARK memory limit assigned to self (set by default or the user through setServiceRelativeMemoryLimitMB
  • If the total use is below the low watermark, no action is taken
  • If we are above the watermark, a fixed number of service instances are selected for pause

To pause a service, the following criteria have to be met:

  • The service has not seen any updates in the past few maintenance intervals, thus deemed inactive
  • The service is in AVAILABLE stage
  • The service has the ON_DEMAND_LOAD option set

An eligible service will then be:

  • removed from the active dispatching map
  • set to ProcessingStage.PAUSED
  • Its service object instance will be serialized to bytes, using Utils.toBytes()
  • a POST will be sent to the /core/service-context-index ephemeral index store (so it will persisted to disk, and associated with a self link)

Resume

Service pause is logically transparent to an external client (service or user). The only noticeable side effects are increased latency to complete the first operation received after a pause, and some additional per instance stats that document the number of pause/resume actions.

A service is resumed, transparently, when an operation is received by the service host. The host will

  • Query the service context blob index, using the self link as the primary key
  • If a blob, representing the serialized service exists, the service is deserialized into a java runtime object
  • The service is re-attached to the active dispatch map and set to ProcessingStage.AVAILABLE
  • The inbound operation that caused the resume is processed

Validation

A long running test in our CI proves that a Xenon service host process, limited to 64MB (or some other small number) can support millions of service instances, created at the rate of 1M per hour, and still run with over 80% of heap available. In addition, it randomly picks "paused" services and issues requests, to verify resume works

Clone this wiki locally