Memcached Ticket Registry

Memcached integration is enabled by including the following dependency in the WAR overlay:


MemCacheTicketRegistry stores tickets in one or more memcached instances. The spymemcached library used by this component presents memcached as a key/value store that accepts String keys and Java Object values. Memcached stores data in exactly one node among many in a distributed cache, thus avoiding the requirement to replicate or otherwise share data between nodes. A deterministic function is used to locate the node, N’, on which to store key K:

N' = f(h(K), N1, N2, N3, ... Nm)

where h(K) is the hash of key K, N1 … Nm is the set of cache nodes, and N’N … Nm.

The function is deterministic in that it consistently produces the same result for a given key and set of cache nodes. Note that a change in the set of available cache nodes may produce a different target node on which to store the key.

Configuration Considerations

There are three core configuration concerns with memcached:

  1. Hash Algorithm
  2. Node locator strategy
  3. Object serialization mechanism

Hash Algorithm

The hash algorithm is used to transform a key value into a memcached storage key that uniquely identifies the corresponding value. The choice of hashing algorithm has implications for failover behavior that is important for HA deployments. The FNV1_64_HASH algorithm is recommended since it offers a nice balance of speed and low collision rate; see the javadocs for alternatives.

Node Locator

The node locator serves as the deterministic node selection function for the memcached client provided by the underlying spymemcached library. There are two choices:


The array modulus mechanism is the default and suitable for cases when the number of nodes in the memcached pool is expected to be consistent. The algorithm simply computes an index into the array of memcached nodes:

hash(key) % length(nodes)

Obviously the selected index is a function of the number of memcached nodes, so variance in number of nodes produces variance in the node selected to store the key, which is undesirable.

The consistent strategy generally provides a target node that does not vary with the number of nodes. This strategy should be used in cases where the memcached pool may grow or shrink dynamically, including due to frequent node failure.

Object Serialization

Memcached stores bytes of data, so CAS tickets must be serialized to a byte array prior to storage. CAS ships with a custom serialization component KryoTranscoder based on the Kryo serialization framework. This component is recommended over the default Java serialization mechanism since it produces much more compact data, which benefits both storage requirements and throughput.


To see the relevant list of CAS properties, please review this guide.

High Availability Considerations

Memcached does not provide for replication by design, but the client is tolerant to node failures with failureMode="Redistribute". In this mode a write failure will simply cause the client to flag the node as failed and remove it from the set of available nodes. It subsequently recomputes the node location function with the reduced node set to find a new node on which to store the key. If the node location function selects the same node, which is likely for the CONSISTENT strategy, a backup node will be computed. The value is written to and read from the failover node until the primary node recovers. The client will periodically check the failed node for liveliness and restore it to the node pool as soon as it recovers. When the primary node is resurrected, if it contains a value for a particular key, it would supersede the value known to the failover node. The most common effect on CAS behavior in this circumstance would occur when ticket-granting tickets have duplicate values, which could affect single sign-out and prevent access to services. In particular, services accessed and forced authentications that occur while the failover service is active would be lost when the failed node recovers. In most cases this behavior is tolerable, but it can be avoided by restarting the memcached service on the failed node prior to rejoining the cache pool.

A read failure in Redistribute mode causes the node to be removed from the set of available nodes, a failover node is computed, and a value is read from that node. In most cases this results in a key not found situation. The effect on CAS behavior depends on the type of ticket requested:

  • Service ticket - Service access would be denied for the requested ticket, but permitted for subsequent attempts since a new ticket would be generated and validated.
  • Ticket-granting ticket - The SSO session would be terminated and re-authentication would be required.

Read failures are thus entirely innocuous for environments where re-authentication is acceptable.