Call us: +1-415-738-4000

Pinning, Expiration, and Eviction

Introduction

This page covers managing the life of the data in each of BigMemory's data storage tiers, including the pinning features of Automatic Resource Control (ARC).

The following are options for data life within the BigMemory tiers:

  • Flush – To move an entry to a lower tier. Flushing is used to free up resources while still keeping data in BigMemory.
  • Fault – To copy an entry from a lower tier to a higher tier. Faulting occurs when data is required at a higher tier but is not resident there. The entry is not deleted from the lower tiers after being faulted.
  • Eviction – To remove an entry from BigMemory. The entry is deleted; it can only be reloaded from an outside source. Entries are evicted to free up resources.
  • Expiration – A status based on Time To Live and Time To Idle settings. To maintain performance, expired entries may not be immediately flushed or evicted.
  • Pinning – To force data to remain in certain tiers. Pinning can be set on individual entries or an entire cache, and must be used with caution to avoid exhausting a resource such as heap.

Setting Expiration

BigMemory data entries expire based on parameters with configurable values. When eviction occurs, expired elements are the first to be removed. Having an effective expiration configuration is critical to optimizing the use of resources such as heap and maintaining overall performance.

To add expiration, edit the values of the following <cache> attributes, and tune these values based on results of performance tests:

  • timeToIdleSeconds – The maximum number of seconds an element can exist in the BigMemory data store without being accessed. The element expires at this limit and will no longer be returned from BigMemory. The default value is 0, which means no TTI eviction takes place (infinite lifetime).
  • timeToLiveSeconds – The maximum number of seconds an element can exist in the the BigMemory data store regardless of use. The element expires at this limit and will no longer be returned from BigMemory. The default value is 0, which means no TTL eviction takes place (infinite lifetime).
  • eternal – If the cache's eternal flag is set, it overrides any finite TTI/TTL values that have been set. Individual cache elements may also be set to eternal. Eternal elements and caches do not expire, however they may be evicted.

See How Configuration Affects Element Eviction for more information on how configuration can impact eviction.

Pinning Data

Data that should remain in the data store regardless of resource constraints can be pinned. You can pin individual entries, or an entire cache.

Pinning Individual Cache Entries

Some APIs like OpenJPA and Hibernate require pinning of specific Elements. Specific entries can be programmatically pinned to the containing cache:

cache.setPinned(key, true);

The entry can be unpinned by the same method:

cache.setPinned(key, false);

To unpin all of a cache's pinned entries:

cache.unpinAll();

To check if an entry is pinned:

cache.isPinned(key); // Returns a boolean: true if the key is pinned.

Pinning a cache entry guarantees its storage in memory (heap or off-heap).

Note that pinning is a status applied to a cache entry's key. The entry's value may be null, and any operations on the value have no effect on the pinning status.

Pinning a Cache

Entire caches can be pinned using the pinning element in cache configuration. This element has a required attribute (store) to specify which data tiers the cache should be pinned to:

<pinning store="localMemory" />

The store attribute can have one of the following values:

  • localHeap – Cache data is pinned to the memory store (on heap). Unexpired entries can never be flushed to a lower tier or be evicted.
  • localMemory – Cache data is pinned to the memory store or the off-heap store. Unexpired entries can never be flushed to a lower tier or be evicted.
  • inCache – Cache data is pinned in the cache, which can be in any tier cache data is stored. The tier is chosen based on performance-enhancing efficiency algorithms.

For example, the following cache is configured to pin its entries:

<cache name="Cache1" ... >
     <pinning store="inCache" />
</cache>

The following cache is configured to pin its entries to heap or off-heap only:

<cache name="Cache2" ... >
     <pinning store="localMemory" />
</cache>

If only a limited set of a cache's entries should be pinned, it may be more efficient to pin those individual elements rather than the entire cache.

Scope of Pinning

Pinning achieved programmatically will not be persisted — after a restart the pinned entries are no longer pinned. Cache pinning in configuration is reinstated with the configuration file.

How Configuration Affects Element Flushing and Eviction

The following example shows a cache with certain expiration settings:

<cache name="myCache"
      eternal="false" timeToIdleSeconds="3600"
      timeToLiveSeconds="0" memoryStoreEvictionPolicy="LFU">
</cache>

Note the following about the myCache configuration:

  • If a client accesses an entry in myCache that has been idle for more than an hour (timeToIdleSeconds), that element is evicted.
  • If an entry expires but is not accessed, and no resource constraints force eviction, then the expired entry remains in place.
  • Entries in myCache can live forever if accessed at least once per 60 minutes (timeToLiveSeconds). However, unexpired entries may still be flushed based on other limitations (see Sizing BigMemory Tiers).

Pinning Overrides Cache Sizing

Pinning takes priority over configured cache sizes. For example, in the following cache the pinning configuration overrides the maxEntriesOnHeap setting:

<cache name="myCache"
     maxEntriesOnHeap="10000"
     maxBytesLocalOffHeap="8g"
     ... >
    <pinning store="localHeap" />
</cache>

While expired cache entries (even ones that have been pinned) can always be flushed and eventually evicted, most non-expired elements can be flushed and evicted as well, if resource limitations are reached. However, pinned elements, whether pinned individually or resident in a pinned cache, cannot be evicted if they haven't expired.

CAUTION: Pinning Could Cause Failures
Potentially, pinned caches could grow to an unlimited size. Caches should never be pinned unless they are designed to hold a limited amount of data (such as reference data) or their usage and expiration characteristics are understood well enough to conclude that they cannot cause errors.

Data Freshness and Expiration

Databases and other Systems of Record (SOR) that were not built to accommodate caching outside of the database do not normally come with any default mechanism for notifying external processes when data has been updated or modified.

When using BigMemory as a caching system, the following strategies can help to keep the data in the cache in sync:

  • Data Expiration: Use the eviction algorithms included with Ehcache, along with the timeToIdleSeconds and timetoLiveSeconds settings, to enforce a maximum time for elements to live in the cache (forcing a re-load from the database or SOR).

  • Message Bus: Use an application to make all updates to the database. When updates are made, post a message onto a message queue with a key to the item that was updated. All application instances can subscribe to the message bus and receive messages about data that is updated, and can synchronize their local copy of the data accordingly (for example by invalidating the cache entry for updated data)

  • Triggers: Using a database trigger can accomplish a similar task as the message bus approach. Use the database trigger to execute code that can publish a message to a message bus. The advantage to this approach is that updates to the database do not have to be made only through a special application. The downside is that not all database triggers support full execution environments and it is often unadvisable to execute heavy-weight processing such as publishing messages on a queue during a database trigger.

The Data Expiration method is the simplest and most straightforward. It gives you the most control over the data synchronization, and doesn't require cooperation from any external systems. You simply set a data expiration policy and let Ehcache expire data from the cache, thus allowing fresh reads to re-populate and re-synchronize the cache.

If you choose the Data Expiration method, you can read more about the cache configuration settings at cache eviction algorithms, and review the timeToIdle and timeToLive configuration settings. The most important consideration when using the expiration method is balancing data freshness with database load. The shorter you make the expiration settings - meaning the more "fresh" you try to make the data - the more load you will incur on the database.

Try out some numbers and see what kind of load your application generates. Even modestly short values such as 5 or 10 minutes will afford significant load reductions.