Stream-Relational Processing Platforms

In past posts I had discussed the differences between Spark and Flink in terms of stream processing, with Spark treating streaming data as micro-batches, and Flink treating streaming data as a first-class citizen.  Apache Beam is another streaming framework that has an interesting goal of generalizing other stream processing platforms.

Recently I joined Confluent, so I was interested in learning how Beam differs from Kafka Streams.  Kafka Streams is a stream processing framework on top of Kafka that has both a functional DSL as well as a declarative SQL layer called KSQL.  Confluent has an informative blog that compares Kafka Streams with Beam in great technical detail.  From what I’ve gathered, one way to state the differences between the two systems is as follows:

  • Kafka Streams is a stream-relational processing platform.
  • Apache Beam is a stream-only processing platform.

A stream-relational processing platform has the following capabilities which are typically missing in a stream-only processing platform:

  • Relations (or tables) are first-class citizens, i.e. each has an independent identity.
  • Relations can be transformed into other relations.
  • Relations can be queried in an ad-hoc manner.

For example, in Beam one can aggregate state using windowing, but this state cannot be queried in an ad-hoc manner.  It can only be emitted back into the stream through the use of triggers.  This is why Beam can be viewed as a stream-only system.

The capability to transform one relation into another is what gives relational algebra (and SQL) its power.  Operations on relations such as selection, projection, and Cartesian product result in more relations, which can then be used for further operations.  This is referred to as the closure property of relational algebra.

Ad-hoc querying allows the user to inject himself as a “sink” at various points in the stream.  At these junctures, he can construct interactive queries to extract the information that he is really interested in.  The user’s temporal presence becomes intertwined with the stream.  This essentially gives the stream a more dynamic nature, something that can help the larger organization be more dynamic as well.

Jay Kreps has written a lot about stream-table (or stream-relation) duality, where streams and tables (or relations) can be considered as two ways of looking at the same data.  This duality has also been explored in research projects like CQL, which was one of the first systems to provide support for both streams and relations1.   In the CQL paper, the authors discuss how both stream-relational systems and stream-only systems are equally expressive.  However, the authors also point out the ease-of-use of stream-relational systems over stream-only systems:

…the dual approach results in more intuitive queries than the stream-only approach.2

Modern day stream processing platforms attempt to unify the world of batch and stream processing.  Batches are essentially treated as bounded streams.  However, stream-relational processing platforms go beyond stream-only processing platforms by not only unifying batch and stream processing, but also unifying the world of streams and relations.  This means that streams can be converted to relations and vice versa within the data flow of the framework itself (and not just at the endpoints of the flow).

Beam is meant to generalize stream-only systems.  Streaming engines can plug into Beam by implementing a Beam Runner.  In order to understand both Kafka Streams and Beam, I spent a couple of weekends writing a Beam Runner for Kafka Streams3.  It provides a proof-of-concept of how the Beam Dataflow model can be implemented on top of Kafka Streams.  However, it doesn’t make use of the ability of Kafka Streams to represent relations, nor does it use the native windowing functionality of Kafka Streams, so it is not as useful to someone who might want the full power of stream-relational functionality4.

I would like to dwell a bit more on stream-table duality as I feel it is a fundamental concept in computer science, not just in databases.  As Jay Kreps has mentioned, stream-table duality has been explored extensively by CQL (2003).  However, a similar duality was known by the functional programming community even earlier.  Here is an interesting reference from over 30 years ago:

“Now we have seen that streams provide an alternative way to model objects with local state.  We can model a changing quantity, such as the local state of some object, using a stream that represents the time history of successive states.  In essence, we represent time explicitly, using streams, so that we decouple time in our simulated world from the sequence of events that take place during evaluation.”5

The above quote is from Structure and Interpretation of Computer Programs, a classic text in computer science.  From a programming perspective, modeling objects with streams leads to a functional programming view of time, whereas modeling objects with state leads to an imperative programming view of time.  So perhaps stream-table duality can also be called stream-state duality.  Streams and state are alternative ways of modeling the temporality of objects in the external world.  This is one of the key ideas behind event sourcing.

If one wants to get even more abstract, analogous concepts exist in philosophy, specifically within the metaphysics of time.  For example,

“…eternalism and presentism are conflicting views about the nature of time.  The eternalist claims that all times exist and thus that the objects present at all past, present, and future times exist.  In contrast, the presentist argues that only the present exists, and thus that only those objects present now exist.  As a result the eternalist, but not the presentist, can sincerely quantify over all times and objects existing at those times.”6

So a Beam proponent would be close to an eternalist, while a MySQL advocate would be close to a presentist.  With Kafka Streams, you get to be both!

Now that I’ve been able to encompass all of temporal reality with this blog, let me presently return to the main point.  Kafka Streams, as a stream-relational processing platform, provides first-class support for relations, relation transformations, and ad-hoc querying, which are all missing in a generalized stream-only framework like Beam7 .  And as the world of relational databases has shown (for almost 50 years since Codd’s original paper), the use of relations, along with being able to transform and query them in an interactive fashion, can provide a critically important tool to enterprises that wish to be more responsive amidst changing business needs and circumstances.






Stream-Relational Processing Platforms

Graph Analytics on HBase with HGraphDB and Apache Flink Gelly

Previously, I’ve shown how both Apache Giraph and Apache Spark GraphFrames can be used to analyze graphs stored in HGraphDB.  In this blog I will show how yet another graph analytics framework, Apache Flink Gelly, can be used with HGraphDB.

First, some observations on how Giraph, GraphFrames, and Gelly differ.  Giraph runs on Hadoop MapReduce, while GraphFrames and Gelly run on Spark and Flink, respectively. MapReduce has been pivotal in launching the era of big data.  Two of its characteristics are the following:

  1. MapReduce has only 3 steps (the map, shuffle, and reduce steps)
  2. MapReduce processes data in batches

The fact that MapReduce utilizes only 3 steps has led to the development of workflow engines like Apache Oozie that can combine MapReduce jobs into more complex flows, represented by directed acyclic graphs.  Also, the fact that MapReduce performs only batch processing has led to the development of stream processing frameworks like Apache Storm, which is often combined with Hadoop MapReduce in what is referred to as a lambda architecture.

Later, dataflow engines such as Apache Spark and Apache Flink were developed to handle data processing as a single job, rather than several independent MapReduce jobs that need to be chained together.  However, while Spark is fundamentally a batch-oriented framework, Flink is fundamentally a stream-oriented framework.   Both try to unify batch and stream processing in different ways.  Spark provides stream processing by breaking data into micro-batches.  Flink posits that batch is a special case of streaming, and that stream-processing engines can handle batches better than batch-processing engines can handle streams.

Needless to say, users who have the requirement to process big data, including large graphs, have a plethora of unique and interesting options at their disposal today.

To use Apache Flink Gelly with HGraphDB, graph data first needs to be wrapped in Flink DataSets.  HGraphDB provides two classes, HBaseVertexInputFormat and HBaseEdgeInputFormat, than can be used to import the vertices and edges of a graph into DataSets.

As a demonstration, we can run one of the Gelly neighborhood examples on HGraphDB as follows.  First we create the graph in the example:

Vertex v1 = graph.addVertex(, 1L);
Vertex v2 = graph.addVertex(, 2L);
Vertex v3 = graph.addVertex(, 3L);
Vertex v4 = graph.addVertex(, 4L);
Vertex v5 = graph.addVertex(, 5L);
v1.addEdge("e", v2, "weight", 0.1);
v1.addEdge("e", v3, "weight", 0.5);
v1.addEdge("e", v4, "weight", 0.4);
v2.addEdge("e", v4, "weight", 0.7);
v2.addEdge("e", v5, "weight", 0.3);
v3.addEdge("e", v4, "weight", 0.2);
v4.addEdge("e", v5, "weight", 0.9);

A vertex in Gelly consists of an ID and a value, whereas an edge in Gelly consists of the source vertex ID, the target vertex ID, and an optional value.  When using HBaseVertexInputFormat and HBaseEdgeInputFormat, the name of a property can be specified for the property value in the HGraphDB vertex or edge to be associated with the Gelly vertex or edge.  If no property name is specified, then the value will default to the ID of the vertex or edge.  Below we import the vertices using an instance of HBaseVertexInputFormat with no property name specified, and we import the edges using an instance of HBaseEdgeInputFormat with the property name specified as “weight”.

HBaseGraphConfiguration conf = graph.configuration();
ExecutionEnvironment env = 
DataSet<Tuple2<Long, Long>> vertices = env.createInput(
    new HBaseVertexInputFormat<>(conf),
    TypeInformation.of(new TypeHint<Tuple2<Long, Long>>() {})
DataSet<Tuple3<Long, Long, Double>> edges = env.createInput(
    new HBaseEdgeInputFormat<>(conf, "weight"),
    TypeInformation.of(new TypeHint<Tuple3<Long, Long, Double>>() {})

Once we have the two DataSets, we can create a Gelly graph as follows:

Graph<Long, Long, Double> gelly = 
    Graph.fromTupleDataSet(vertices, edges, env);

Finally, running the neighborhood processing example is exactly the same as in the documentation:

DataSet<Tuple2<Long, Double>> minWeights = 
    gelly.reduceOnEdges(new SelectMinWeight(), EdgeDirection.OUT);

// user-defined function to select the minimum weight
static final class SelectMinWeight implements ReduceEdgesFunction {
    public Double reduceEdges(
        Double firstEdgeValue, Double secondEdgeValue) {
        return Math.min(firstEdgeValue, secondEdgeValue);

HGraphDB brings together several big data technologies in the Apache ecosystem in order to process large graphs. Graph data can be stored in Apache HBase, OLTP graph operations can be performed using Apache TinkerPop, and complex graph analytics can be performed using Apache Giraph, Apache Spark GraphFrames, or Apache Flink Gelly.

Graph Analytics on HBase with HGraphDB and Apache Flink Gelly

HBase Application Archetypes Redux

At Yammer, we’ve transitioned away from polyglot persistence to persistence consolidation. In a microservice architecture, the principle that each microservice should be responsible for its own data had led to a proliferation of different types of data stores at Yammer. This in turn led to multiple efforts to make sure that each data store could be easily used, monitored, operationalized, and maintained. In the end, we decided it would be more efficient, both architecturally and organizationally, to reduce the number of data store types in use at Yammer to as few as possible.

Today HBase is the primary data store for non-relational data at Yammer (we use PostgreSQL for relational data).  Microservices are still responsible for their own data, but the data is segregated by cluster boundaries or mechanisms within the data store itself (such as HBase namespaces or PostgreSQL schemas).

HBase was chosen for a number of reasons, including its performance, scalability, reliability, its support for strong consistency, and its ability to support a wide variety of data models.  At Yammer we have a number of services that rely on HBase for persistence in production:

  • Feedie, a feeds service
  • RoyalMail, an inbox service
  • Ocular, for tracking messages that a user has viewed
  • Streamie, for storing activity streams
  • Prankie, a ranking service with time-based decay
  • Authlog, for authorization audit trails
  • Spammie, for spam monitoring and blocking
  • Graphene, a generic graph modeling service

HBase is able to satisfy the persistence needs of several very different domains. Of course, there are some use cases for which HBase is not recommended, for example, when using raw HDFS would be more efficient, or when ad-hoc querying via SQL is preferred (although projects like Apache Phoenix can provide SQL on top of HBase).

Previously, Lars George and Jonathan Hsieh from Cloudera attempted to survey the most commonly occurring use cases for HBase, which they referred to as application archetypes.  In their presentation, they categorized archetypes as either “good”, “bad”, or “maybe” when used with HBase. Below I present an augmented listing of their “good” archetypes, along with pointers to projects that implement them.


The Entity archetype is the most natural of the archetypes.  HBase, being a wide column store, can represent the entity properties with individual columns.  Projects like Apache Gora and HEntityDB support this archetype.

Column Family: default
Row Key Column: <property 1 name> Column: <property 2 name>
<entity ID>  <property 1 value> <property 2 value>

Entities can be also stored in the same manner as with a key-value store.  In this case the entity would be serialized as a binary or JSON value in a single column.

Column Family: default
Row Key Column: body
<entity ID>  <entity blob>


The Sorted Collection archetype is a generalization of the original Messaging archetype that was presented.  In this archetype the entities are stored as binary or JSON values, with the column qualifier being the value of the sort key to use.  For example, in a messaging feed, the column qualifier would be a timestamp or a monotonically increasing counter of some sort.  The column qualifier can also be “inverted” (such as by subtracting a numeric ID from the maximum possible value) so that entities are stored in descending order.

Column Family: default
Row Key Column: <sort key 1 value> Column: <sort key 2 value>
<collection ID>  <entity 1 blob> <entity 2 blob>

Alternatively, each entity can be stored as a set of properties.  This is similar to how Cassandra implements CQL.  HEntityDB supports storing entity collections in this manner.

Column Family: default
Row Key Column: <sort key 1 value + property 1 name> Column: <sort key 1 value + property 2 name> Column: <sort key 2 value + property 1 name> Column: <sort key 2 value + property 2 name>
<collection ID> <property 1 of entity 1> <property 2 of entity 1> <property 1 of entity 2> <property 2 of entity 2>

In order to access entities by some other value than the sort key, additional column families representing indices can be used.

Column Family: sorted Column Family: index
Row Key Column: <sort key 1 value> Column: <sort key 2 value> Column: <index 1 value> Column: <index 2 value>
<collection ID>  <entity 1 blob> <entity 2 blob> <entity 1 blob> <entity 2 blob>

To prevent the collection from growing unbounded, a coprocessor can be used to trim the sorted collection during compactions.  If index column families are used, the coprocessor would also remove corresponding entries from the index column families when trimming the sorted collection.  At Yammer, both the Feedie and RoyalMail services use this technique.  Both services also use server-side filters for efficient pagination of the sorted collection during queries.


Using a technique called key-flattening, a document can be shredded by storing each value in the document according to the path from the root to the name of the element containing the value.  HDocDB uses this approach.

Column Family: default
Row Key Column: <property 1 path> Column: <property 2 path>
<document ID>  <property 1 value> <property 2 value>

The document can also be stored as a binary value, in which case support for Medium Objects (MOBs) can be used if the documents are large.  This approach is described in the book Architecting HBase Applications.

Column Family: default
Row Key Column: body
<document ID>  <reference to MOB>


There are many ways to store a graph in HBase.  One method is to use an adjacency list, where each vertex stores its neighbors in the same row.  This is the approach taken in JanusGraph.

Column Family: default
Row Key Column: <edge 1 key> Column: <edge 2 key> Column: <property 1 name> Column: <property 2 name>
<vertex ID>  <edge 1 properties> <edge 2 properties> <property 1 value> <property 2 value>

In the table above, the edge key is actually comprised of a number of parts, including the label, direction, edge ID, and adjacent vertex ID.

Alternatively, a separate table to represent edges can be used, in which case the incident vertices are stored in the same row as an edge.   This may scale better if the adjacency list is large, such as in a social network.  This is the approach taken in both Zen and HGraphDB.

Column Family: default
Row Key Column: <property 1 name> Column: <property 2 name>
<vertex ID> <property 1 value> <property 2 value>
Column Family: default
Row Key Column: fromVertex Column: toVertex Column: <property 1 name> Column: <property 2 name>
<edge ID>  <vertex ID> <vertex ID> <property 1 value> <property 2 value>

When storing edges in a separate table, additional index tables must be used to provide efficient access to the incident edges of a vertex.  For example, the full list of tables in HGraphDB can be viewed here.


A queue can be modeled by using a row key comprised of the consumer ID and a counter.  Both Cask and Box implement queues in this manner.

Column Family: default
Row Key Column: metadata Column: body
<consumer ID + counter>  <message metadata> <message body>

Cask also uses coprocessors for efficient scan filtering and queue trimming, and Apache Tephra for transactional queue processing.


The Metrics archetype is a variant of the Entity archetype in which the column values are counters or some other aggregate.

Column Family: default
Row Key Column: <property 1 name> Column: <property 2 name>
<entity ID>  <property 1 counter> <property 2 counter>

HGraphDB is actually a combination of the Graph and Metrics archetypes, as arbitrary counters can be stored on either vertices or edges.

Update: For other projects that use HBase, see this list.

HBase Application Archetypes Redux

Graph Analytics on HBase with HGraphDB and Spark GraphFrames

In a previous post, I showed how to analyze graphs stored in HGraphDB using Apache Giraph.  Giraph depends on Hadoop, and some developers may be using Spark instead.  In this blog I will show how to analyze HGraphDB graphs using Apache Spark GraphFrames.

In order to prepare data stored in HGraphDB for GraphFrames, we need to import vertex and edge data from HGraphDB into Spark DataFrames.  Hortonworks provides a Spark-on-HBase Connector to do just that.  The Spark-on-HBase Connector allows for custom serde (serializer/deserializer) types to be created by implementing the SHCDataType trait.  The serde for HGraphDB is available here.  (When testing the serde, I ran into some issues with the Spark-on-HBase Connector for which I have submitted pull requests.  Hopefully those will be merged soon.  In the meantime, you can use my fork of the Spark-on-HBase Connector.  Update:  Thanks to HortonWorks, these have been merged.)

To demonstrate how to use HGraphDB with GraphFrames, we first use HGraphDB to create the same graph example that is used in the GraphFrames User Guide.

Vertex a = graph.addVertex(, "a", "name", "Alice", "age", 34);
Vertex b = graph.addVertex(, "b", "name", "Bob", "age", 36);
Vertex c = graph.addVertex(, "c", "name", "Charlie", "age", 30);
Vertex d = graph.addVertex(, "d", "name", "David", "age", 29);
Vertex e = graph.addVertex(, "e", "name", "Esther", "age", 32);
Vertex f = graph.addVertex(, "f", "name", "Fanny", "age", 36);
Vertex g = graph.addVertex(, "g", "name", "Gabby", "age", 60);
a.addEdge("friend", b);
b.addEdge("follow", c);
c.addEdge("follow", b);
f.addEdge("follow", c);
e.addEdge("follow", f);
e.addEdge("friend", d);
d.addEdge("friend", a);
a.addEdge("friend", e);

Now that the graph is stored in HGraphDB, we need to specify a schema to be used by the Spark-on-HBase Connector for retrieving vertex and edge data.

def vertexCatalog = s"""{
    |"table":{"namespace":"testGraph", "name":"vertices",
    |  "tableCoder":"org.apache.spark.sql.execution.datasources.hbase.types.HGraphDB", "version":"2.0"},
      |"id":{"cf":"rowkey", "col":"key", "type":"string"},
      |"name":{"cf":"f", "col":"name", "type":"string"},
      |"age":{"cf":"f", "col":"age", "type":"int"}

def edgeCatalog = s"""{
    |"table":{"namespace":"testGraph", "name":"edges",
    |  "tableCoder":"org.apache.spark.sql.execution.datasources.hbase.types.HGraphDB", "version":"2.0"},
      |"id":{"cf":"rowkey", "col":"key", "type":"string"},
      |"relationship":{"cf":"f", "col":"~l", "type":"string"},
      |"src":{"cf":"f", "col":"~f", "type":"string"},
      |"dst":{"cf":"f", "col":"~t", "type":"string"}

Some things to note about this schema:

  • The HGraphDB serde is specified as the tableCoder above.
  • All HGraphDB columns are stored in a column family named f.
  • Vertex and edge labels are stored in a column with qualifier ~l.
  • The source and destination columns have qualifiers ~f and ~t, respectively.
  • All vertex and edge properties are stored in columns with the qualifiers simply being the name of the property.

Now that we have a schema, we can create Spark DataFrames for both the vertices and edges, and then pass these to the GraphFrame factory.

def withCatalog(cat: String): DataFrame = {
val verticesDataFrame = withCatalog(vertexCatalog)
val edgesDataFrame = withCatalog(edgeCatalog)
val g = GraphFrame(verticesDataFrame, edgesDataFrame)

With the GraphFrame in hand, we now have full access to the Spark GraphFrame APIs. For instance, here are some arbitrary graph operations from the GraphFrames Quick Start.

// Query: Get in-degree of each vertex.

// Query: Count the number of "follow" connections in the graph.
g.edges.filter("relationship = 'follow'").count()

// Run PageRank algorithm, and show results.
val results = g.pageRank.resetProbability(0.01).maxIter(20).run()"id", "pagerank").show()

You can see further graph operations against our example graph (taken from the GraphFrames User Guide) in this test.

As you can see, HGraphDB makes graphs stored in HBase easily accessible by Apache TinkerPopApache Giraph, and now Apache Spark GraphFrames.

Graph Analytics on HBase with HGraphDB and Spark GraphFrames

Don’t Settle For Eventual Consistency

This week Google released Cloud Spanner1, a publicly available version of their Spanner database. This completes the public release of their 3 main databases, Bigtable (released as Cloud Bigtable), Megastore (released as Cloud Datastore), and Spanner. Spanner is the culmination of Google’s research in data stores, which provides a globally distributed, relational database that is both strongly consistent and highly available.

But doesn’t the CAP theorem state that we have to choose consistency over availability, or availability over consistency? Over the years, Google has been arguing that you can have both strong consistency and high availability, and that you don’t have to settle for eventual consistency. In fact, all 3 of Google’s data stores are strongly consistent systems.

Some Background

In 2000, Brewer came up with the CAP conjecture2, which was later proved as a theorem by Gilbert and Lynch3. It states that you can choose only 2 of the 3 properties:

  • C: consistency (or linearizability)
  • A: 100% availability (in the context of network partitions)
  • P: tolerance of network partitions

Later Coda Hale made the point that you can’t sacrifice partition tolerance, so really the choice is between CP and AP (and not CA)4.

What is the tradeoff?

According to the CAP theorem, when you choose a data store, you must choose either an AP system (that is eventually consistent) or a CP system (that is strongly consistent). But Google would argue the following points:

  1. In AP systems, client code becomes more complex and error-prone in order to deal with inconsistencies.
  2. AP systems are not 100% available in practice.
  3. CP systems can be made to be highly available in practice.
  4. From the above 3 points, when you choose availability over consistency, you are not gaining 100% availability but you are losing consistency and you are gaining complexity.

Let’s drill down into these points.

Client complexity

Here is what Google has to say about using AP systems:

“We also have a lot of experience with eventual consistency systems at Google. In all such systems, we find developers spend a significant fraction of their time building extremely complex and error-prone mechanisms to cope with eventual consistency and handle data that may be out of date. We think this is an unacceptable burden to place on developers and that consistency problems should be solved at the database level.”5

This has led Google to focus on data stores that are CP.

AP systems in practice

Many engineers are confused about the definition of “availability” in the CAP theorem. Most engineers think of availability in terms of a service level agreement (SLA) or a service level objective (SLO), which is typically measured in “9s”. However, as Kleppmann has pointed out, the “availability” in the CAP theorem is not a measurement or a metric, but a liveness property of an algorithm.6 I am going to distinguish between the two types of availability by referring to them as “effective availability” and “algorithmic availability”.

  • Effective availability: the empirically measured percentage of successful requests over some period, often measured in “9s”.
  • Algorithmic availability: a liveness property of an algorithm where every request to a non-failing node must eventually return a valid response.

The CAP theorem is only concerned with algorithmic availability.  An algorithmic availability of 100% does not guarantee an effective availability of 100%. The algorithmic availability from the CAP theorem only applies if both the implementation and the execution of the algorithm is without error. In practice, most outages to an AP system are not due to network issues, which the algorithm can handle, but rather to implementation defects, user errors, misconfiguration, resource limits, and misbehaving clients. Google found that in Spanner only 7.6% of its errors were network-related, whereas 52.5% of errors were user-related (such as overload and misconfiguration) and 13.3% of errors were due to bugs. Google actually refers to these errors as “incidents” since they were able to prevent most of them from affecting availability.7

At Yammer we have experience with AP systems, and we’ve seen loss of availability for both Cassandra and Riak for various reasons.  Our AP systems have not been more reliable than our CP systems, yet they have been more difficult to work with and reason about in the presence of inconsistencies.  Other companies have also seen outages with AP systems in production.8 So in practice, AP systems are just as susceptible as CP systems to outages due to issues such as human error and buggy code, both on the client side and the server side.

CP systems in practice

With Spanner, Google is able to attain an availability of 5 “9s”, which is 5.26 minutes of downtime per year.7 Likewise, Facebook uses HBase, another CP system based on Bigtable, and claims to be able to attain an availability of between 4 to 5 “9s”.9 In practice, mature CP systems can be made to be highly available. In fact, due to its strong consistency and high availability, Google refers to Spanner as “effectively” CA, which means they are focusing on effective availability (a practical measure) and not algorithmic availability (a theoretical property).

A bad tradeoff?

With an AP system, you are giving up consistency, and not really gaining anything in terms of effective availability, the type of availability you really care about.  Some might think you can regain strong consistency in an AP system by using strict quorums (where the number of nodes written + number of nodes read > number of replicas).  Cassandra calls this “tunable consistency”.  However, Kleppmann has shown that even with strict quorums, inconsistencies can result.10  So when choosing (algorithmic) availability over consistency, you are giving up consistency for not much in return, as well as gaining complexity in your clients when they have to deal with inconsistencies.


There’s nothing wrong with using an AP system in general. An AP system might exhibit the lower latencies that you require (such as with a cache), or perhaps your data is immutable so you don’t care as much about strong consistency, or perhaps 99.9% consistency is “good enough”.11 These are all valid reasons for accepting eventual consistency.  However, in practice AP systems are not necessarily more highly available than CP systems, so don’t settle for eventual consistency in order to gain availability. The availability you think you will be getting (effective) is not the availability you will actually get (algorithmic), which will not be as useful as you might think.






  1. D. Srivastava. Introducing Cloud Spanner: a global database service for mission-critical applications, 2017 
  2. E. Brewer. Towards robust distributed systems. Proceedings of the 19th Annual ACM Symposium on Principles of Distributed Computing, Portland, OR, 2000 
  3. S. Gilbert, N. Lynch. Brewer’s conjecture and the feasibility of consistent, available, partition-tolerant web services. ACM SIGACT News 33(2), 2002 
  4. C. Hale. You Can’t Sacrifice Partition Tolerance, 2010 
  5.  J. Corbett, J. Dean, M. Epstein, A. Fikes, C. Frost, JJ Furman, S. Ghemawat, A. Gubarev, C. Heiser, P. Hochschild, W. Hsieh, S. Kanthak, E. Kogan, H. Li, A. Lloyd, S. Melnik, D. Mwaura, D. Nagle, S. Quinlan, R. Rao, L. Rolig, Y. Saito, M. Szymaniak, C. Taylor, R. Wang, and D. Woodford. Spanner: Google’s Globally-Distributed Database. Proceedings of OSDI ‘12: Tenth Symposium on Operating System Design and Implementation, Hollywood, CA, October, 2012 
  6. M. Kleppmann. A Critique of the CAP Theorem, 2015 
  7. E. Brewer. Spanner, TrueTime, and the CAP Theorem, 2017 
  8. D. Nadolny. PagerDuty: One Year of Cassandra Failures, 2015 
  9. Z. Fong, R. Shroff. HydraBase – The evolution of HBase@Facebook, 2014 
  10. M. Kleppmann. Designing Data-Intensive Applications, Chapter 9, p 334, 2017 
  11. P. Bailis, A. Ghodsi. Eventual consistency today: limitations, extensions, and beyond. Commun. ACM 56(5), 55–63, 2013 
Don’t Settle For Eventual Consistency

Graph Analytics on HBase with HGraphDB and Giraph

HGraphDB is a client framework for HBase that provides a TinkerPop Graph API.  HGraphDB also provides integration with Apache Giraph, a graph compute engine for analyzing graphs that Facebook has shown to be massively scalable.  In this blog we will show how to convert a sample Giraph computation that works with text files to instead work with HGraphDB.

In the Giraph quick start, the SimpleShortestPathsComputation is used to show how to run a Giraph computation against a graph contained in a file as a JSON representation.  Here are the contents of the JSON file:


Each line above has the format [fromVertexId, vertexValue, [[toVertexId, edgeValue],...]], where the edgeValue is the weight or cost of the edge that will be used for the path computation.

To run the example in the Giraph quick start, the following command line is used:

hadoop jar giraph-examples-1.3.0-SNAPSHOT-for-hadoop-2.5.1-jar-with-dependencies.jar \
    org.apache.giraph.GiraphRunner \
    org.apache.giraph.examples.SimpleShortestPathsComputation \
    -vif \
    -vip /user/ryokota/input/tiny_graph.txt \
    -vof \
    -op /user/ryokota/output/shortestpaths \
    -w 1 -ca giraph.SplitMasterWorker=false

The results of the job will appear in a file under the output path (/user/ryokota/output/shortestpaths), with the following contents:

0 1.0
1 0.0
2 2.0
3 1.0
4 5.0

Now let’s leave that example and consider the exact same graph stored in HGraphDB.  The graph above can be created in HGraphDB using the following statements.

        Vertex v0 = graph.addVertex(, 0);
        Vertex v1 = graph.addVertex(, 1);
        Vertex v2 = graph.addVertex(, 2);
        Vertex v3 = graph.addVertex(, 3);
        Vertex v4 = graph.addVertex(, 4);
        v0.addEdge("e", v1, "weight", 1);
        v0.addEdge("e", v3, "weight", 3);
        v1.addEdge("e", v0, "weight", 1);
        v1.addEdge("e", v2, "weight", 2);
        v1.addEdge("e", v3, "weight", 1);
        v2.addEdge("e", v1, "weight", 2);
        v2.addEdge("e", v4, "weight", 4);
        v3.addEdge("e", v0, "weight", 3);
        v3.addEdge("e", v1, "weight", 1);
        v3.addEdge("e", v4, "weight", 4);
        v4.addEdge("e", v3, "weight", 4);
        v4.addEdge("e", v2, "weight", 4);

There is also a class called HBaseBulkLoader that can be used for more efficient creation of larger graphs.

Instead of using the JSON input format above, HGraphDB provides two input formats, HBaseVertexInputFormat and HBaseEdgeInputFormat, which will read from the vertices table and edges table in HBase, respectively.  To use these formats, the Giraph computation needs to be changed slightly.  Here is the original SimpleShortestPathsComputation:

public class SimpleShortestPathsComputation extends BasicComputation<LongWritable, DoubleWritable, FloatWritable, DoubleWritable> {
  public void compute(
      Vertex<LongWritable, DoubleWritable, FloatWritable> vertex,
      Iterable<DoubleWritable> messages) throws IOException {
    if (getSuperstep() == 0) {
      vertex.setValue(new DoubleWritable(Double.MAX_VALUE));
    double minDist = isSource(vertex) ? 0d : Double.MAX_VALUE;
    for (DoubleWritable message : messages) {
      minDist = Math.min(minDist, message.get());
    if (minDist < vertex.getValue().get()) {
      vertex.setValue(new DoubleWritable(minDist));
      for (Edge<LongWritable, FloatWritable> edge : vertex.getEdges()) {
        double distance = minDist + edge.getValue().get();
        sendMessage(edge.getTargetVertexId(), new DoubleWritable(distance));

And here is the version for HGraphDB.  The main changes are in bold.

public class SimpleShortestPathsComputation extends
        HBaseComputation<Long, DoubleWritable, FloatWritable, DoubleWritable> {
  public void compute(
      Vertex<ObjectWritable<Long>, VertexValueWritable<DoubleWritable>, EdgeValueWritable<FloatWritable>> vertex,
      Iterable<DoubleWritable> messages) throws IOException {
    VertexValueWritable<DoubleWritable> vertexValue = vertex.getValue();
    if (getSuperstep() == 0) {
      vertexValue.setValue(new DoubleWritable(Double.MAX_VALUE));
    double minDist = isSource(vertex) ? 0d : Double.MAX_VALUE;
    for (DoubleWritable message : messages) {
      minDist = Math.min(minDist, message.get());
    if (minDist < vertexValue.getValue().get()) {
      vertexValue.setValue(new DoubleWritable(minDist));
      for (Edge<ObjectWritable, EdgeValueWritable> edge : vertex.getEdges()) {
        double distance = minDist + ((Number) edge.getValue().getEdge().property("weight").value()).doubleValue();
        sendMessage(edge.getTargetVertexId(), new DoubleWritable(distance));

The major difference is that when using HBaseVertexInputFormat, the “value” of a Giraph vertex is an instance of type VertexValueWritable, which is comprised of an HBaseVertex and a Writable value.   Likewise when using HBaseEdgeInputFormat, the “value” of a Giraph edge is an instance of type EdgeValueWritable, which is comprised of an HBaseEdge and a Writable value.  The instances of HBaseVertex and HBaseEdge should be considered read-only and only be used to obtain IDs and property values.

Running the above Giraph computation against HBase is similar to running the original example.  Note that we also have to customize IdWithValueTextOutputFormat to work properly with VertexValueWritable.

./hadoop jar hgraphdb-0.4.4-SNAPSHOT-test-jar-with-dependencies.jar \
    org.apache.giraph.GiraphRunner \
    io.hgraphdb.giraph.examples.SimpleShortestPathsComputation \
    -vif io.hgraphdb.giraph.HBaseVertexInputFormat \
    -eif io.hgraphdb.giraph.HBaseEdgeInputFormat \
    -vof io.hgraphdb.giraph.examples.IdWithValueTextOutputFormat \
    -op /user/ryokota/output/shortestpaths \
    -w 1 -ca giraph.SplitMasterWorker=false \
    -ca hbase.zookeeper.quorum= \
    -ca zookeeper.znode.parent=/hbase-unsecure \
    -ca gremlin.hbase.namespace=testgraph \
    -ca hbase.mapreduce.edgetable=testgraph:edges \
    -ca hbase.mapreduce.vertextable=testgraph:vertices

As an alternative to using a text-based output format such as IdWithValueTextOutputFormat, HGraphDB provides two abstract output formats, HBaseVertexOutputFormat and HBaseEdgeOutputFormat, that can be used to modify the graph after a Giraph computation.  For example, the shortest path result for each vertex could be set as a property on the vertex by extending HBaseVertexOutputFormat and implementing the method

public abstract void writeVertex(HBaseBulkLoader writer, HBaseVertex vertex, Writable value);

As you can see, HGraphDB extends the functionality in Apache Giraph by making it quite easy to both read and write graphs stored in HBase when performing sophisticated graph analytics.

Graph Analytics on HBase with HGraphDB and Giraph

HGraphDB: HBase as a TinkerPop Graph Database

The use of graph databases is common among social networking companies. A social network can easily be represented as a graph model, so a graph database is a natural fit. For instance, Facebook has a graph database called Tao, Twitter has FlockDB, and Pinterest has Zen. At Yammer, an enterprise social network, we rely on HBase for much of our messaging infrastructure, so I decided to see if HBase could also be used for some graph modelling and analysis.

Below I put together a wish list of what I wanted to see in a graph database.

  • It should be implemented directly on top of HBase.
  • It should support the TinkerPop 3 API.
  • It should allow the user to supply IDs for both vertices and edges.
  • It should allow user-supplied IDs to be either strings or numbers.
  • It should allow property values to be of arbitrary type, including maps, arrays, and serializable objects.
  • It should support indexing vertices by label and property.
  • It should support indexing edges by label and property, specific to a given vertex.
  • It should support range queries and pagination with both vertex indices and edge indices.

I did not find a graph database that met all of the above criteria. For instance, Titan is a graph database that supports the TinkerPop API, but it is not implemented directly on HBase. Rather, it is implemented on top of an abstraction layer that can be integrated with HBase, Cassandra, or Berkeley DB as its underlying store. Also, Titan does not support user-supplied IDs. S2Graph is a graph database that is implemented directly on HBase, and it supports both user-supplied IDs and indices on edges, but it does not yet support the TinkerPop API nor does it support indices on vertices.

This led me to create HGraphDB, a TinkerPop 3 layer for HBase. It provides support for all of the above bullet points. Feel free to try it out if you are interested in using HBase as a graph database.

HGraphDB: HBase as a TinkerPop Graph Database