Chapter 10. Titan Cache

10.1. Caching

Titan employs multiple layers of data caching to facilitate fast graph traversals. The caching layers are listed here in the order they are accessed from within a Titan transaction. The closer the cache is to the transaction, the faster the cache access and the higher the memory footprint and maintenance overhead.

10.2. Transaction-Level Caching

Within an open transaction, Titan maintains two caches:

  • Vertex Cache: Caches accessed vertices and their adjacency list (or subsets thereof) so that subsequent access is significantly faster within the same transaction. Hence, this cache speeds up iterative traversals.
  • Index Cache: Caches the results for index queries so that subsequent index calls can be served from memory instead of calling the index backend and (usually) waiting for one or more network round trips.

The size of both of those is determined by the transaction cache size. The transaction cache size can be configured via cache.tx-cache-size or on a per transaction basis by opening a transaction via the transaction builder graph.buildTransction() and using the setVertexCacheSize(int) method.

10.2.1. Vertex Cache

The vertex cache contains vertices and the subset of their adjacency list that has been retrieved in a particular transaction. The maximum number of vertices maintained in this cache is equal to the transaction cache size. If the transaction workload is an iterative traversal, the vertex cache will significantly speed it up. If the same vertex is not accessed again in the transaction, the transaction level cache will make no difference.

Note, that the size of the vertex cache on heap is not only determined by the number of vertices it may hold but also by the size of their adjacency list. In other words, vertices with large adjacency lists (i.e. many incident edges) will consume more space in this cache than those with smaller lists.

Furthermore note, that modified vertices are pinned in the cache, which means they cannot be evicted since that would entail loosing their changes. Therefore, transaction which contain a lot of modifications may end up with a larger than configured vertex cache.

10.2.2. Index Cache

The index cache contains the results of index queries executed in the context of this transaction. Subsequent identical index calls will be served from this cache and are therefore significantly cheaper. If the same index call never occurs twice in the same transaction, the index cache makes no difference.

Each entry in the index cache is given a weight equal to 2 + result set size and the total weight of the cache will not exceed half of the transaction cache size.

10.3. Database Level Caching

The database level cache retains adjacency lists (or subsets thereof) across multiple transactions and beyond the duration of a single transaction. The database level cache is shared by all transactions across a database. It is more space efficient than the transaction level caches but also slightly slower to access. In contrast to the transaction level caches, the database level caches do not expire immediately after closing a transaction. Hence, the database level cache significantly speeds up graph traversals for read heavy workloads across transactions.

Chapter 12, Configuration Reference lists all of the configuration options that pertain to Titan’s database level cache. This page attempts to explain their usage.

Most importantly, the database level cache is disabled by default in the current release version of Titan. To enable it, set cache.db-cache=true.

10.3.1. Cache Expiration Time

The most important setting for performance and query behavior is the cache expiration time which is configured via cache.db-cache-time. The cache will hold graph elements for at most that many milliseconds. If an element expires, the data will be re-read from the storage backend on the next access.

If there is only one Titan instance accessing the storage backend or if this instance is the only one modifying the graph, the cache expiration can be set to 0 which disables cache expiration. This allows the cache to hold elements indefinitely (unless they are evicted due to space constraints or on update) which provides the best cache performance. Since no other Titan instance is modifying the graph, there is no danger of holding on to stale data.

If there are multiple Titan instances accessing the storage backend, the time should be set to the maximum time that can be allowed between another Titan instance modifying the graph and this Titan instance seeing the data. If any change should be immediately visible to all Titan instances, the database level cache should be disabled in a distributed setup. However, for most applications it is acceptable that a particular Titan instance sees remote modifications with some delay. The larger the maximally allowed delay, the better the cache performance. Note, that a given Titan instance will always immediately see its own modifications to the graph irrespective of the configured cache expiration time.

10.3.2. Cache Size

The configuration option cache.db-cache-size controls how much heap space Titan’s database level cache is allowed to consume. The larger the cache, the more effective it will be. However, large cache sizes can lead to excessive GC and poor performance.

The cache size can be configured as a percentage (expressed as a decimal between 0 and 1) of the total heap space available to the JVM running Titan or as an absolute number of bytes.

Note, that the cache size refers to the amount of heap space that is exclusively occupied by the cache. Titan’s other data structures and each open transaction will occupy additional heap space. If additional software layers are running in the same JVM, those may occupy a significant amount of heap space as well (e.g. Rexster, embedded Cassandra, etc). Be conservative in your heap memory estimation. Configuring a cache that is too large can lead to out-of-memory exceptions and excessive GC.

10.3.3. Clean Up Wait Time

When a vertex is locally modified (e.g. an edge is added) all of the vertex’s related database level cache entries are marked as expired and eventually evicted. This will cause Titan to refresh the vertex’s data from the storage backend on the next access and re-populate the cache.

However, when the storage backend is eventually consistent, the modifications that triggered the eviction may not yet be visible. By configuring cache.db-cache-clean-wait, the cache will wait for at least this many milliseconds before repopulating the cache with the entry retrieved from the storage backend.

If Titan runs locally or against a storage backend that guarantees immediate visibility of modifications, this value can be set to 0.

10.4. Storage Backend Caching

Each storage backend maintains its own data caching layer. These caches benefit from compression, data compactness, coordinated expiration and are often maintained off heap which means that large caches can be used without running into garbage collection issues. While these caches can be significantly larger than the database level cache, they are also slower to access.

The exact type of caching and its properties depends on the particular storage backend. Please refer to the respective documentation for more information about the caching infrastructure and how to optimize it.