High-performance distributed analytical database + Spark SQL queries + built for streaming.

filodb-announce google group

    _______ __      ____  ____ 
   / ____(_) /___  / __ \/ __ )
  / /_  / / / __ \/ / / / __  |
 / __/ / / / /_/ / /_/ / /_/ / 
/_/   /_/_/\____/_____/_____/  

Columnar, versioned layers of data wrapped in a yummy high-performance analytical database engine.

See architecture and datasets and reading for more information. Also note the Spark Notebooks under doc... there is one for time-series/geo analysis of the NYC Taxi dataset!

Table of Contents generated with DocToc

• Overview
  • Use Cases
  • Anti-use-cases
  • Roadmap
• Pre-requisites
• Getting Started
• Introduction to FiloDB Data Modelling
  • Computed Columns
  • FiloDB vs Cassandra Data Modelling
  • Data Modelling and Performance Considerations
  • Predicate Pushdowns
  • Example FiloDB Schema for machine metrics
  • Distributed Partitioning
• Using FiloDB Data Source with Spark
  • Configuring FiloDB
  • Spark Data Source API Example (spark-shell)
  • Spark/Scala/Java API
  • Spark Streaming Example
  • SQL/Hive Example
  • Querying Datasets
• Using the CLI
  • Running the CLI
  • CLI Example
• Current Status
• Deploying
• Code Walkthrough
• Building and Testing
• You can help!

FiloDB is a new open-source distributed, versioned, and columnar analytical database designed for modern streaming workloads.

• High performance - competitive with Parquet scan speeds, plus filtering along two or more dimensions
  • Very flexible filtering: filter on only part of a partition key, much more flexible than allowed in Cassandra
• Compact storage - within 35% of Parquet for CassandraColumnStore
  • See the blog post on Apache Cassandra for analytics: a performance and storage analysis
• Idempotent writes - primary-key based appends and updates; easy exactly-once ingestion from streaming sources
• Distributed - pluggable storage engine includes Apache Cassandra and in-memory
• Low-latency - minimal SQL query latency of 15ms on one node; sub-second easily achievable with filtering and easy to use concurrency control
  • See post on 700 SQL Queries per Second in Apache Spark with FiloDB
• SQL queries - plug in Tableau or any tool using JDBC/ODBC drivers
• Ingest from Spark/Spark Streaming from any supported Spark data source

Overview presentation -- see the docs folder for design docs.

To compile the .mermaid source files to .png's, install the Mermaid CLI.

• Storage and analysis of streaming event / time series data
• Data warehousing
• In-memory database for Spark Streaming analytics
• Low-latency in-memory SQL database engine

• Heavily transactional, update-oriented workflows

Your input is appreciated!

• Productionization and automated stress testing
• Kafka input API / connector (without needing Spark)
• In-memory caching for significant query speedup
• True columnar querying and execution, using late materialization and vectorization techniques. GPU/SIMD.
• Projections. Often-repeated queries can be sped up significantly with projections.

1. Java 8
2. SBT to build
3. Apache Cassandra (We prefer using CCM for local testing) (Optional if you are using the in-memory column store)
4. Apache Spark (1.5.x)

1. Clone the project and cd into the project directory,

   $ git clone https://github.com/tuplejump/FiloDB.git
   $ cd FiloDB
   

2. Choose either the Cassandra column store (default) or the in-memory column store.

   • Start a Cassandra Cluster. If its not accessible at localhost:9042 update it in core/src/main/resources/application.conf.
   • Or, use FiloDB's in-memory column store with Spark (does not work with CLI). Pass the --conf spark.filodb.store=in-memory to spark-submit / spark-shell. This is a great option to test things out, and is really really fast!

3. For Cassandra, run filo-cli --command init to initialize the default filodb keyspace.

4. Dataset creation can be done using filo-cli or using Spark Shell / Scala/Java API.

5. Inserting data can be done using filo-cli (CSV only), using Spark SQL/JDBC (INSERT INTO), or the Spark Shell / Scala / Java API.

6. Querying is done using Spark SQL/JDBC or Scala/Java API.

7. Listing/deleting/maintenance can be done using filo-cli. If using Cassandra, cqlsh can also be used to inspect metadata.

Note: There is at least one release out now, tagged via Git and also located in the "Releases" tab on Github.

Perhaps it's easiest by starting with a diagram of how FiloDB stores data.

Three types of key define the data model of a FiloDB table.

1. partition key - decides how data is going to be distributed across the cluster. All data within one partition key is guaranteed to fit on one node. May consist of multiple columns.
2. segment key - groups row values into efficient chunks. Segments within a partition are sorted by segment key and range scans can be done over segment keys. Ideal is > 1000 rows per segment.
3. row key - acts as a primary key within each partition and decides how data will be sorted within each segment. May consist of multiple columns.

The PRIMARY KEY for FiloDB consists of (partition key, row key). When choosing the above values you must make sure the combination of the two are unique. No component of a primary key may be null - see the :getOrElse function for a way of dealing with null inputs.

Specifying the partitioning column is optional. If a partitioning column is not specified, FiloDB will create a default one with a fixed value, which means everything will be thrown into one node, and is only suitable for small amounts of data. If you don't specify a partitioning column, then you have to make sure your row keys are all unique.

For examples of data modeling and choosing keys, see the examples below as well as datasets.

You may specify a function, or computed column, for use with any key column. This is especially useful for working around the non-null requirement for keys, or for computing a good segment key.

• Like Cassandra, partitions (physical rows) distribute data and clustering keys act as a primary key within a partition
• Like Cassandra, a single partition is the smallest unit of parallelism when querying from Spark
• Wider rows work better for FiloDB (bigger chunk/segment size)
• FiloDB does not have Cassandra's restrictions for partition key filtering. You can filter by any partition keys with most operators. This means less tables in FiloDB can match more query patterns.
• Cassandra range scans over clustering keys is available over the segment key

Choosing Partition Keys.

• Partition keys are the most efficient way to filter data. Remember that, unlike Cassandra, FiloDB is able to efficiently filter any column in a partition key -- even string contains, IN on only one column. It can do this because FiloDB pre-scans a much smaller table ahead of scanning the main columnar chunk table. This flexibility means that there is no need to populate different tables with different orders of partition keys just to optimize for different queries.
• If there are too few partitions, then FiloDB will not be able to distribute and parallelize reads.
• If the numer of rows in each partition is too few, then the storage will not be efficient.
• If the partition key is time based, there may be a hotspot in the cluster as recent data is all written into the same set of replicas, and likewise for read patterns as well.

Segment Key and Chunk Size.

Within each partition, data is delineated by the segment key into segments. Within each segment, successive flushes of the MemTable writes data in chunks. The segmentation is key to sorting and filtering data within partitions, and the chunk size (which depends on segmentation) also affects performance. The smaller the chunk size, the higher the overhead of scanning data becomes.

Segmentation and chunk size distribution may be checked by the CLI analyze command. In addition, the following configuration affects segmentation and chunk size:

• memtable.max-rows-per-table, memtable.flush-trigger-rows affects how many rows are kept in the MemTable at a time, and this along with how many partitions are in the MemTable directly leads to the chunk size upon flushing.
• The segment size is directly controlled by the segment key. Choosing a segment key that groups data into big enough chunks (at least 1000 is a good guide) is highly recommended. Experimentation along with running filo-cli analyze is recommended to come up with a good segment key. See the Spark ingestion of GDELT below on an example... choosing an inappropriate segment key leads to MUCH slower ingest and read performance.
• chunk_size option when creating a dataset caps the size of a single chunk.

To help with planning, here is an exact list of the predicate pushdowns (in Spark) that help with reducing I/O and query times:

• Partition key column(s): =, IN on any partition key column
• Segment key: must be of the form segmentKey >/>= value AND segmentKey </<= value

This is one way I would recommend setting things up to take advantage of FiloDB.

The metric names are the column names. This lets you select on just one metric and effectively take advantage of columnar layout.

• Partition key = hostname
• Row key = timestamp (say millis)
• Segment key = :round timestamp 10000 (Timestamp rounded to nearest 10000 millis or 10 seconds)
• Columns: hostname, timestamp, CPU, load_avg, disk_usage, etc.

You can add more metrics/columns over time, but storing each metric in its own column is FAR FAR more efficient, at least in FiloDB. For example, disk usage metrics are likely to have very different numbers than load_avg, and so Filo can optimize the storage of each one independently. Right now I would store them as ints and longs if possible.

With the above layout, as long as there aren’t too many hostnames, set the memtable max size and flush trigger to both high numbers, you should get good read performance. Queries that would work well for the above layout:

• SELECT avg(load_avg), min(load_avg), max(load_avg) FROM metrics WHERE timestamp > t1 AND timestamp < t2 etc.

Queries that would work well once we expose a local Cassandra query interface:

• Select metrics from one individual host

Another possible layout is something like this:

• Partition key = hostname % 1024 (or pick your # of shards)
• Row key = hostname, timestamp

Currently, FiloDB is a library in Spark and requires the user to distribute data such that no two nodes have rows with the same partition key.

• The easiest strategy to accomplish this is to have data partitioned via a queue such as Kafka. That way, when the data comes into Spark Streaming, it is already partitioned correctly.
• Another way of accomplishing this is to use a DataFrame's sort method before using the DataFrame write API.

FiloDB has a Spark data-source module - filodb.spark. So, you can use the Spark Dataframes read and write APIs with FiloDB. To use it, follow the steps below

1. Start Cassandra and update project configuration if required.
2. From the FiloDB project directory, execute,

   $ sbt clean
   $ ./filo-cli --command init
   $ sbt spark/assembly
   

3. Use the jar FiloDB/spark/target/scala-2.10/filodb-spark-assembly-0.2-SNAPSHOT.jar with Spark 1.5.x.

The options to use with the data-source api are:

Partitioning columns could be created using an expression on the original column in Spark:

val newDF = df.withColumn("partition", df("someCol") % 100)

or even UDFs:

val idHash = sqlContext.udf.register("hashCode", (s: String) => s.hashCode())
val newDF = df.withColumn("partition", idHash(df("id")) % 100) 

However, note that the above methods will lead to a physical column being created, so use of computed columns is probably preferable.

Some options must be configured before starting the Spark Shell or Spark application. There are two methods:

1. Modify the application.conf and rebuild, or repackage a new configuration.
2. Override the built in defaults by setting SparkConf configuration values, preceding the filo settings with spark.filodb. For example, to change the keyspace, pass --conf spark.filodb.cassandra.keyspace=mykeyspace to Spark Shell/Submit. To use the fast in-memory column store instead of Cassandra, pass --conf spark.filodb.store=in-memory.
3. It might be easier to pass in an entire configuration file to FiloDB. Pass the java option -Dconfig.file=/path/to/my-filodb.conf, for example using --java-driver-options.

You can follow along using the Spark Notebook... launch the notebook using EXTRA_CLASSPATH=$FILO_JAR ADD_JARS=$FILO_JAR ./bin/spark-notebook & where FILO_JAR is the path to filodb-spark-assembly jar. See the FiloDB notebook to follow the GDELT examples below, or the NYC Taxi notebook for some really neat time series/geo analysis!

Or you can start a spark-shell locally,

bin/spark-shell --jars ../FiloDB/spark/target/scala-2.10/filodb-spark-assembly-0.2-SNAPSHOT.jar --packages com.databricks:spark-csv_2.10:1.2.0 --driver-memory 3G --executor-memory 3G

Loading CSV file from Spark:

scala> val csvDF = sqlContext.read.format("com.databricks.spark.csv").
           option("header", "true").option("inferSchema", "true").
           load("../FiloDB/GDELT-1979-1984-100000.csv")

Creating a dataset from a Spark DataFrame,

scala> import org.apache.spark.sql.SaveMode
import org.apache.spark.sql.SaveMode

scala> csvDF.write.format("filodb.spark").
             option("dataset", "gdelt").
             option("row_keys", "GLOBALEVENTID").
             option("segment_key", ":round GLOBALEVENTID 10000").
             option("partition_keys", ":getOrElse MonthYear -1").
             mode(SaveMode.Overwrite).save()

Above, we partition the GDELT dataset by MonthYear, creating roughly 72 partitions for 1979-1984, with the unique GLOBALEVENTID used as a row key. We group every 10000 eventIDs into a segment using the convenient :round computed column (GLOBALEVENTID is correlated with time, so in this case we could pack segments with consecutive EVENTIDs). You could use multiple columns for the partition or row keys, of course. For example, to partition by country code and year instead:

scala> csvDF.write.format("filodb.spark").
             option("dataset", "gdelt_by_country_year").
             option("row_keys", "GLOBALEVENTID").
             option("segment_key", ":string 0").
             option("partition_keys", ":getOrElse Actor2CountryCode NONE,:getOrElse Year -1").
             mode(SaveMode.Overwrite).save()

Note that in the above case, since events are spread over a much larger number of partitions, it no longer makes sense to use GLOBALEVENTID as a segment key - at least with the original 10000 as a rounding factor. There are very few events for a given country and year within the space of 10000 event IDs, leading to inefficient storage. Instead, we use a single segment for each partition. We probably could have used :round GLOBALEVENTID 500000 or some other bigger factor as well. Using :round GLOBALEVENT 10000 lead to 3x slower ingest and at least 5x slower reads.

The key definitions can be left out for appends:

sourceDataFrame.write.format("filodb.spark").
                option("dataset", "gdelt").
                mode(SaveMode.Append).save()

Note that for efficient columnar encoding, wide rows with fewer partition keys are better for performance.

Reading the dataset,

val df = sqlContext.read.format("filodb.spark").option("dataset", "gdelt").load()

The dataset can be queried using the DataFrame DSL. See the section Querying Datasets for examples.

There is a more typesafe API than the Spark Data Source API.

import filodb.spark._
sqlContext.saveAsFilo(df, "gdelt",
                      rowKeys = Seq("GLOBALEVENTID"),
                      segmentKey = ":round GLOBALEVENTID 10000",
                      partitionKeys = Seq(":getOrElse MonthYear -1"))

The above creates the gdelt table based on the keys above, and also inserts data from the dataframe df.

There is also an API purely for inserting data... after all, specifying the keys is not needed when inserting into an existing table.

import filodb.spark._
sqlContext.insertIntoFilo(df, "gdelt")

The API for creating a DataFrame is also much more concise:

val df = sqlContext.filoDataset("gdelt")

The above method calls rely on an implicit conversion. From Java, you would need to create a new FiloContext first:

FiloContext fc = new filodb.spark.FiloContext(sqlContext);
fc.insertIntoFilo(df, "gdelt");

It's not difficult to ingest data into FiloDB using Spark Streaming. Simple use foreachRDD on your DStream and then transform each RDD into a DataFrame.

For an example, see the StreamingTest.

Start Spark-SQL:

  bin/spark-sql --jars path/to/FiloDB/spark/target/scala-2.10/filodb-spark-assembly-0.2-SNAPSHOT.jar

(NOTE: if you want to connect with a real Hive Metastore, you should probably instead start the thrift server, also adding the --jars above, and then start the spark-beeline client)

Create a temporary table using an existing dataset,

  create temporary table gdelt
  using filodb.spark
  options (
   dataset "gdelt"
 );

Then, start running SQL queries!

You probably want to create a permanent Hive Metastore entry so you don't have to run create temporary table every single time at startup:

  CREATE TABLE gdelt using filodb.spark options (dataset "gdelt");

Once this is done, you could insert data using SQL syntax:

  INSERT INTO TABLE gdelt SELECT * FROM othertable;

Of course, this assumes othertable has a similar schema.

Now do some queries, using the DataFrame DSL:

scala> df.select(count(df("MonthYear"))).show()
...<skipping lots of logging>...
COUNT(MonthYear)
4037998

or SQL, to find the top 15 events with the highest tone:

scala> df.registerTempTable("gdelt")

scala> sqlContext.sql("SELECT Actor1Name, Actor2Name, AvgTone FROM gdelt ORDER BY AvgTone DESC LIMIT 15").collect()
res13: Array[org.apache.spark.sql.Row] = Array([208077.29634561483])

Now, how about something uniquely Spark .. feed SQL query results to MLLib to compute a correlation:

scala> import org.apache.spark.mllib.stat.Statistics

scala> val numMentions = df.select("NumMentions").map(row => row.getInt(0).toDouble)
numMentions: org.apache.spark.rdd.RDD[Double] = MapPartitionsRDD[100] at map at DataFrame.scala:848

scala> val numArticles = df.select("NumArticles").map(row => row.getInt(0).toDouble)
numArticles: org.apache.spark.rdd.RDD[Double] = MapPartitionsRDD[104] at map at DataFrame.scala:848

scala> val correlation = Statistics.corr(numMentions, numArticles, "pearson")

The filo-cli accepts arguments as key-value pairs. The following keys are supported:

You may want to customize a configuration to point at your Cassandra cluster, or change other configuration parameters. The easiest is to pass in a customized config file:

./filo-cli -Dconfig.file=/path/to/myfilo.conf --command init

You may also set the FILO_CONFIG_FILE environment var instead, but any -Dconfig.file args passed in takes precedence.

Individual configuration params may also be changed by passing them on the command line. They must be the first arguments passed in. For example:

./filo-cli -Dfilodb.columnstore.segment-cache-size=10000 --command ingestcsv ....

The following examples use the GDELT public dataset and can be run from the project directory.

Create a dataset with all the columns :

./filo-cli --command create --dataset gdelt --columns GLOBALEVENTID:int,SQLDATE:string,MonthYear:int,Year:int,FractionDate:double,Actor1Code:string,Actor1Name:string,Actor1CountryCode:string,Actor1KnownGroupCode:string,Actor1EthnicCode:string,Actor1Religion1Code:string,Actor1Religion2Code:string,Actor1Type1Code:string,Actor1Type2Code:string,Actor1Type3Code:string,Actor2Code:string,Actor2Name:string,Actor2CountryCode:string,Actor2KnownGroupCode:string,Actor2EthnicCode:string,Actor2Religion1Code:string,Actor2Religion2Code:string,Actor2Type1Code:string,Actor2Type2Code:string,Actor2Type3Code:string,IsRootEvent:int,EventCode:string,EventBaseCode:string,EventRootCode:string,QuadClass:int,GoldsteinScale:double,NumMentions:int,NumSources:int,NumArticles:int,AvgTone:double,Actor1Geo_Type:int,Actor1Geo_FullName:string,Actor1Geo_CountryCode:string,Actor1Geo_ADM1Code:string,Actor1Geo_Lat:double,Actor1Geo_Long:double,Actor1Geo_FeatureID:int,Actor2Geo_Type:int,Actor2Geo_FullName:string,Actor2Geo_CountryCode:string,Actor2Geo_ADM1Code:string,Actor2Geo_Lat:double,Actor2Geo_Long:double,Actor2Geo_FeatureID:int,ActionGeo_Type:int,ActionGeo_FullName:string,ActionGeo_CountryCode:string,ActionGeo_ADM1Code:string,ActionGeo_Lat:double,ActionGeo_Long:double,ActionGeo_FeatureID:int,DATEADDED:string,Actor1Geo_FullLocation:string,Actor2Geo_FullLocation:string,ActionGeo_FullLocation:string --rowKeys GLOBALEVENTID --segmentKey ':string 0'

Verify the dataset metadata:

./filo-cli --command list --dataset gdelt

Import data from a CSV file:

./filo-cli --command importcsv --dataset gdelt --filename GDELT-1979-1984-100000.csv

Query/export some columns:

./filo-cli --dataset gdelt --select MonthYear,Actor2Code --limit 5 --outfile out.csv

Version 0.2 is the stable, latest released version. It has been tested on a cluster for a different variety of schemas, has a stable data model and ingestion, and features a huge number of improvements over the previous version:

• Multi column partition and row keys
• Separate segment key for much easier control of columnar chunk size and sort order
• Much richer projection key filtering (IN and =, on any column)
• Range scans within partitions by segment key
• Computed columns for easily deriving segment and partition keys, especially useful for time series
• Support for timestamp columns
• Better performance and error handling: true node locality for reads; up to 2x faster scan performance; fast single partition reads; retries and configurable connection timeouts etc.
• A ton of bug fixes
• An experimental feature to sync FiloDB tables to Hive MetaStore

The current version assumes Spark 1.5.x and Cassandra 2.1.x or 2.2.x.

• sbt spark/assembly
• sbt cli/assembly
• Copy core/src/main/resources/application.conf and modify as needed for your own config file
• Set FILO_CONFIG_FILE to the path to your custom config
• Run the cli jar as the filo CLI command line tool and initialize keyspaces if using Cassandra: filo-cli-*.jar --command init

There is a branch for Datastax Enterprise 4.8 / Spark 1.4. Note that if you are using DSE or have vnodes enabled, a lower number of vnodes (16 or less) is STRONGLY recommended as higher numbers of vnodes slows down queries substantially and basically prevents subsecond queries from happening.

Please go to the architecture doc.

Run the tests with sbt test, or for continuous development, sbt ~test. Noisy cassandra logs can be seen in filodb-test.log.

To run benchmarks, from within SBT:

cd jmh
jmh:run -i 5 -wi 5 -f3

You can get the huge variety of JMH options by running jmh:run -help.

• Send your use cases for OLAP on Cassandra and Spark
  • Especially IoT and Geospatial
• Email if you want to contribute

Your feedback will help decide the next batch of features, such as: - which data types to add support for - what architecture is best supported