Eventually-Consistent Storage Backends
When running JanusGraph against an eventually consistent storage backend special JanusGraph features must be used to ensure data consistency and special considerations must be made regarding data degradation.
This page summarizes some of the aspects to consider when running JanusGraph on top of an eventually consistent storage backend like Apache Cassandra or Apache HBase.
On eventually consistent storage backends, JanusGraph must obtain locks
in order to ensure consistency because the underlying storage backend
does not provide transactional isolation. In the interest of efficiency,
JanusGraph does not use locking by default. Hence, the user has to
decide for each schema element that defines a consistency constraint
whether or not to use locking. Use
to explicitly enable locking on a schema element as shown in the
1 2 3 4 5 6
mgmt = graph.openManagement() name = mgmt.makePropertyKey('consistentName').dataType(String.class).make() index = mgmt.buildIndex('byConsistentName', Vertex.class).addKey(name).unique().buildCompositeIndex() mgmt.setConsistency(name, ConsistencyModifier.LOCK) // Ensures only one name per vertex mgmt.setConsistency(index, ConsistencyModifier.LOCK) // Ensures name uniqueness in the graph mgmt.commit()
When updating an element that is guarded by a uniqueness constraint,
JanusGraph uses the following protocol at the end of a transaction when
- Acquire a lock on all elements that have a consistency constraint
- Re-read those elements from the storage backend and verify that they match the state of the element in the current transaction prior to modification. If not, the element was concurrently modified and a PermanentLocking exception is thrown.
- Persist the state of the transaction against the storage backend.
- Release all locks.
This is a brief description of the locking protocol which leaves out optimizations (e.g. local conflict detection) and detection of failure scenarios (e.g. expired locks).
The actual lock application mechanism is abstracted such that JanusGraph can use multiple implementations of a locking provider. Currently, two locking providers are supported in the JanusGraph distribution:
- A locking implementation based on key-consistent read and write operations that is agnostic to the underlying storage backend as long as it supports key-consistent operations (which includes Cassandra and HBase). This is the default implementation and uses timestamp based lock applications to determine which transaction holds the lock.
- A Cassandra specific locking implementation based on the Astyanax locking recipe.
Both locking providers require that clocks are synchronized across all machines in the cluster.
The locking implementation is not robust against all failure scenarios. For instance, when a Cassandra cluster drops below quorum, consistency is no longer ensured. Hence, it is suggested to use locking-based consistency constraints sparingly with eventually consistent storage backends. For use cases that require strict and or frequent consistency constraint enforcement, it is suggested to use a storage backend that provides transactional isolation.
Data Consistency without Locks
Because of the additional steps required to acquire a lock when committing a modifying transaction, locking is a fairly expensive way to ensure consistency and can lead to deadlock when very many concurrent transactions try to modify the same elements in the graph. Hence, locking should be used in situations where consistency is more important than write latency and the number of conflicting transactions is small.
In other situations, it may be better to allow conflicting transactions to proceed and to resolve inconsistencies at read time. This is a design pattern commonly employed in large scale data systems and most effective when the actual likelihood of conflict is small. Hence, write transactions don’t incur additional overhead and any (unlikely) conflict that does occur is detected and resolved at read time and later cleaned up. JanusGraph makes it easy to use this strategy through the following features.
Because edge are stored as single records in the underlying storage
backend, concurrently modifying a single edge would lead to conflict.
Instead of locking, an edge label can be configured to use
ConsistencyModifier.FORK. The following example creates a new edge
related and defines its consistency to FORK.
1 2 3 4
mgmt = graph.openManagement() related = mgmt.makeEdgeLabel('related').make() mgmt.setConsistency(related, ConsistencyModifier.FORK) mgmt.commit()
When modifying an edge whose label is configured to FORK the edge is deleted and the modified edge is added as a new one. Hence, if two concurrent transactions modify the same edge, two modified copies of the edge will exist upon commit which can be resolved during querying traversals if needed.
Edge forking only applies to MULTI edges. Edge labels with a multiplicity constraint cannot use this strategy since a constraint is built into the edge label definition that requires an explicit lock or use the conflict resolution mechanism of the underlying storage backend.
Modifying single valued properties on vertices concurrently can result
in a conflict. Similarly to edges, one can allow an arbitrary number of
properties on a vertex for a particular property key defined with
cardinality LIST and FORK on modification. Hence, instead of conflict
one reads multiple properties. Since JanusGraph allows properties on
properties, provenance information like
author can be added to the
properties to facilitate resolution at read time.
See multi-properties to learn how to define those.
On eventually consistent storage backends, writes may not be immediately visible to the entire cluster causing temporary inconsistencies in the graph. This is an inherent property of eventual consistency, in the sense, that accepted updates must be propagated to other instances in the cluster and no guarantees are made with respect to read atomicity in the interest of performance.
From JanusGraph’s perspective, eventual consistency might cause the following temporary graph inconsistencies in addition the general inconsistency that some parts of a transaction are visible while others aren’t yet.
Stale Index entries
Index entries might point to nonexistent vertices or edges. Similarly, a vertex or edge appears in the graph but is not yet indexed and hence ignored by global graph queries.
Only one direction of an edge gets persisted or deleted which might lead to the edge not being or incorrectly being retrieved.
In order to avoid that write failures result in permanent
inconsistencies in the graph it is recommended to use storage backends
that support batch write atomicity and to ensure that write atomicity
is enabled. To get the benefit of write atomicity, the number
modifications made in a single transaction must be smaller than the
buffer-size option documented in Configuration Reference. The
buffer size defines the maximum number of modifications that
JanusGraph will persist in a single batch. If a transaction has more
modifications, the persistence will be split into multiple batches
which are persisted individually which is useful for batch loading but
invalidates write atomicity.
A permanent inconsistency that can arise when operating JanusGraph on eventually consistent storage backend is the phenomena of ghost vertices. If a vertex gets deleted while it is concurrently being modified, the vertex might re-appear as a ghost.
The following strategies can be used to mitigate this issue:
Configure transactions to (double) check for the existence of vertices prior to returning them. Please see Transaction Configuration for more information and note that this can significantly decrease performance. Note, that this does not fix the inconsistencies but hides some of them from the user.
Run regular batch-jobs to repair inconsistencies in the graph using JanusGraph with TinkerPop’s Hadoop-Gremlin. This is the only strategy that can address all inconsistencies and effectively repair them. We will provide increasing support for such repairs in future versions of Faunus.
Instead of deleting vertices, they are marked as deleted which keeps them in the graph for future analysis but hides them from user-facing transactions.