Ingesting Offline data

Segments for offline tables are constructed outside of Pinot, typically in Hadoop via map-reduce jobs and ingested into Pinot via REST API provided by the Controller. Pinot provides libraries to create Pinot segments out of input files in AVRO, JSON or CSV formats in a hadoop job, and push the constructed segments to the controllers via REST APIs.

When an Offline segment is ingested, the controller looks up the table’s configuration and assigns the segment to the servers that host the table. It may assign multiple servers for each segment depending on the number of replicas configured for that table.

Pinot supports different segment assignment strategies that are optimized for various use cases.

Once segments are assigned, Pinot servers get notified via Helix to “host” the segment. The segments are downloaded from the remote segment store to the local storage, untarred, and memory-mapped.

Once the server has loaded (memory-mapped) the segment, Helix notifies brokers of the availability of these segments. The brokers start to include the new segments for queries. Brokers support different routing strategies depending on the type of table, the segment assignment strategy, and the use case.

Data in offline segments are immutable (Rows cannot be added, deleted, or modified). However, segments may be replaced with modified data.

Starting from release-0.11.0, Pinot supports uploading offline segments to real-time tables. This is useful when user wants to bootstrap a real-time table with some initial data, or add some offline data to a real-time table without changing the data stream. Note that this is different from the setup, and no time boundary is maintained between the offline segments and the real-time segments.

Ingesting Realtime Data

Segments for realtime tables are constructed by Pinot servers with rows ingested from data streams such as Kafka. Rows ingested from streams are made available for query processing as soon as they are ingested, thus enabling applications such as those that need real-time charts on analytics.

In large scale installations, data in streams is typically split across multiple stream partitions. The underlying stream may provide consumer implementations that allow applications to consume data from any subset of partitions, including all partitions (or, just from one partition).

A pinot table can be configured to consume from streams in one of two modes:

• LowLevel: This is the preferred mode of consumption. Pinot creates independent partition-level consumers for each partition. Depending on the the configured number of replicas, multiple consumers may be created for each partition, taking care that no two replicas exist on the same server host. Therefore you need to provision at least as many hosts as the number of replicas configured.

• HighLevel: Pinot creates one stream-level consumer that consumes from all partitions. Each message consumed could be from any of the partitions of the stream. Depending on the configured number of replicas, multiple stream-level consumers are created, taking care that no two replicas exist on the same server host. Therefore you need to provision exactly as many hosts as the number of replicas configured.

Of course, the underlying stream should support either mode of consumption in order for a Pinot table to use that mode. Kafka has support for both of these modes. See for more information on the support of other data streams in Pinot.

In either mode, Pinot servers store the ingested rows in volatile memory until either one of the following conditions are met:

1. A certain number of rows are consumed

2. The consumption has gone on for a certain length of time

(See on how to set these values, or have pinot compute them for you)

Upon reaching either one of these limits, the servers do the following:

• Pause consumption

• Persist the rows consumed so far into non-volatile storage

• Continue consuming new rows into volatile memory again.

The persisted rows form what we call a completed segment (as opposed to a consuming segment that resides in volatile memory).

In LowLevel mode, the completed segments are persisted the into local non-volatile store of pinot server as well as the segment store of the pinot cluster (See ). This allows for easy and automated mechanisms for replacing pinot servers, or expanding capacity, etc. Pinot has that ensure that the completed segment is equivalent across all replicas.

During segment completion, one winner is chosen by the controller from all the replicas as the committer server. The committer server builds the segment and uploads it to the controller. All the other non-committer servers follow one of these two paths:

1. If the in-memory segment is equivalent to the committed segment, the non-committer server also builds the segment locally and replaces the in-memory segment

2. If the in-memory segment is non equivalent to the committed segment, the non-committer server downloads the segment from the controller.

For more details on this protocol, please refer to .

In HighLevel mode, the servers persist the consumed rows into local store (and not the segment store). Since consumption of rows can be from any partition, it is not possible to guarantee equivalence of segments across replicas.

See for details.

Need for special NULL value handling

By default, Pinot transforms null values coming from the data source to a default value determined by the type of the corresponding column (or as specified in the schema). Eg: for INT column, the default will be 0 and for STRING column, the default is "null". This transformation is necessary to ensure all the indices can be built correctly during segment creation. However, we're now unable to keep track of the null values in the Pinot table and hence cannot support queries such as:

There is a workaround by matching with default values in the filter predicate. However, this is error prone since oftentimes it's difficult to distinguish valid values from the default null values. Therefore, we added first class NULL value support in Pinot for overcoming this limitation. As of today, the latest version supports NULL filter predicates only. Generic support for NULL handling in query execution is in progress (eg: within aggregation functions such as count or sum).

High Level Architecture

To turn on NULL handling, simply enable the boolean flag in the table index config called as nullHandlingEnabled (please see ). Please note - this will cause Pinot to use additional memory and disk space per segment. The details are as follows:

Ingestion Phase

During data ingestion (either realtime/offline) eachGenericRow object derived from the original data source record keeps track of all the column names containing null values. This is done as part of the NullValueTransformer. For each such column, the segment creation logic updates a NULL value vector (implemented by a roaring bitmap) with the corresponding document ID. Effectively, at the end of the segment creation process we get a per column NULL value vector which can give us the set of document IDs containing null values for that column. This per column vector is then exposed through the DataSource interface for use in query execution.

Query Phase

During Query execution, if the query includes a IS NULL or IS NOT NULL predicate as shown above, we fetch the NULL value vector for the corresponding column within FilterPlanNode and retrieve the corresponding bitmap which represents all document IDs containing NULL values for that column. This bitmap is then used to create a BitmapBasedFilterOperatorwhich does the actual filtering operation.

Many data analytics use-cases only need aggregated data. For example, data used in charts can be aggregated down to one row per time bucket per dimension combination.

Doing this results in much less storage and better query performance. Configuring this for a table is done via the Aggregation Config in the table config.

Aggregation Config

The aggregation config controls the aggregations that happen during realtime data ingestion. Offline aggregations must be handled separately.

Below is a description of the config, which is defined in the ingestion config of the table config.

Requirements

The following are required for ingestion aggregation to work:

• Ingestion aggregation config is effective only for realtime tables. (There is no ingestion time aggregation support for offline tables. We need use or pre-process aggregations in the offline data flow using batch processing engines like Spark/MapReduce).

• type must be lowLevel.

• All metrics must have aggregation configs.

Example Scenario

Here is an example of sales data, where only the daily sales aggregates per product are needed.

Example Input Data

Schema

Note that the schema only reflects the final table structure.

Table Config

From the below aggregation config example, please note that price exists in the input data while total_sales exists in the Pinot Schema.

Example Final Table

Allowed Aggregation Functions

Frequently Asked Questions

Why not use a Startree?

Startrees can only be added to realtime segments after the segments has sealed, and creating startrees is CPU-intensive. Ingestion Aggregation works for consuming segments and uses no additional CPU.

Startrees take additional memory to store, while ingestion aggregation stores less data than the original dataset.

When to not use ingestion aggregation?

If the original rows in non-aggregated form are needed, then ingestion-aggregation cannot be used.

I already use the aggregateMetrics setting?

The aggregateMetrics works the same as Ingestion Aggregation, but only allows for the SUM function.

The current changes are backward compatible, so no need to change your table config unless you need a different aggregation function.

Does this config work for offline data?

Ingestion Aggregation only works for realtime ingestion. For offline data, the offline process needs to generate the aggregates separately.

Why do all metrics need to be aggregated?

If a metric isn't aggregated then it will result in more than one row per unique set of dimensions.

Overview

The new Pinot query engine version 2 (a.k.a Multi-Stage V2 query engine) is designed to support more complex SQL semantics such as JOIN, OVER window, MATCH_RECOGNIZE and eventually, make Pinot support closer to full ANSI SQL semantics.

It also resolves the bottleneck effect for the broker reduce stage where only a single machine is dedicated to perform heavy lifting such as high cardinality GROUP BY result merging; ORDER BY sorting, etc.

Enable the multi-stage query engine

Check if the multi-stage engine is enabled

To check whether your version of Pinot has enabled the multi-stage engine, run a query via the controller console with the "use V2 Engine" flag checked and you should see results similar to the following page.

Manually enable the multi-stage engine

For some older master release versions of Pinot, the multi-stage engine might not be default enabled. We recommend upgrading to Pinot 1.0.

If upgrading to latest master is not an option and you see the "use V2 Engine" flag option on the controller query console, do the following:

1. Add the following configurations to your cluster config (use the POST API in ):

2. Note: The"pinot.query.server.port" and "pinot.query.runner.port" options should only be added if every Pinot component is launched within its own network environment; otherwise, port conflict will occur.

How to programmatically access the multi-stage query engine

There's 2 main way to query the multi-stage engine:

Via REST APIs

Both the Controller query API and the Broker query API allows optional JSON payload for configuration. For example:

• For Controller REST API

• For Broker REST API

Via Query Options

When executing a query via a non-REST API route. we also support enabling the multi-stage engine via the query option. simply attach a query option at the top of your query will enable the multi-stage engine

Troubleshoot

The V2 query engine is still in the beta phase, there might be various performance or feature gaps from the current query engine.

Here are the general troubleshooting steps:

Cluster errors

If you can't get the cluster to run and received an error indicating the multi-stage engine is not available:

• Make sure your pinot-servers have access to the configured port: The default setting above assumes all pinot-servers are run on individual networks and all of them have access to the same port number available within that environment. If your cluster is not set up this way, you will have to manually configure them directly in your pinot-server config file.

Semantic / Runtime errors

• Try downloading the latest docker image or building from the latest master commit

  • We continuously push bug fixes for the multi-stage engine so bugs you encountered might have already been fixed in the latest master build

• Try rewriting your query

Timeout errors

• Try reducing the size of the table(s) used:

  • Adding higher selectivity filters to the tables

• Try executing part of the subquery or a simplified version of the query first.

How to share feedback

Please report any bugs in Apache Pinot Slack . Please include:

• the table/schema config(s)

• the cluster config (zookeeper config, and each components config and scale)

• the problematic SQL query string and corresponding ERROR messages.

Limitations

We are continuously improving the multi-stage engine. However, since the multi-stage engine is still in beta-testing phase, there are some limitations to call out:

• Incomplete data type support: multi-value columns and some other non-primitive data types are not supported. For example SELECT * with multi-value columns will fail.

• The intermediate stages of the multi-stage engine are running purely on heap memory, thus executing a large table join will cause potential out-of-memory errors

For more up-to-date tracking of feature and performance support please follow the Github tracking issues:

• SQL feature tracker:

• Performance and stability tracker:

Reference: Design Docs

The overall PEP design doc and discussion can be found in the following links

• and

Raw source data often needs to undergo some transformations before it is pushed to Pinot.

Transformations include extracting records from nested objects, applying simple transform functions on certain columns, filtering out unwanted columns, as well as more advanced operations like joining between datasets.

A preprocessing job is usually needed to perform these operations. In streaming data sources you might write a Samza job and create an intermediate topic to store the transformed data.

For simple transformations, this can result in inconsistencies in the batch/stream data source and increase maintenance and operator overhead.

To make things easier, Pinot supports transformations that can be applied via the table config.

Transformation Functions

Pinot supports the following functions:

1. Groovy functions

2. Inbuilt functions

Groovy functions

Groovy functions can be defined using the syntax:

Any valid Groovy expression can be used.

⚠️ Enabling Groovy

Allowing execuatable Groovy in ingestion transformation can be a security vulnerability. If you would like to enable Groovy for ingestion, you can set the following controller config.

controller.disable.ingestion.groovy=false

If not set, Groovy for ingestion transformation is disabled by default.

Inbuilt Pinot functions

There are also several inbuilt functions that can be used directly as ingestion transform functions

DateTime functions

These functions enable time transformations.

toEpochXXX

Converts from epoch milliseconds to a higher granularity.

toEpochXXXRounded

Converts from epoch milliseconds to another granularity, rounding to the nearest rounding bucket. For example, 1588469352000 (2020-05-01 42:29:12) is 26474489 minutesSinceEpoch. `toEpochMinutesRounded(1588469352000) = 26474480 (2020-05-01 42:20:00)

fromEpochXXX

Converts from an epoch granularity to milliseconds.

Simple date format

Converts simple date format strings to milliseconds and vice-a-versa, as per the provided pattern string.

JSON functions

Types of transformation

Filtering

Records can be filtered as they are being ingested. A filter function can be specified in the filterConfigs in the ingestionConfigs of the table config.

If the expression evaluates to true, the record will be filtered out. The expressions can use any of the transform functions described in the previous section.

Consider a table that has a column timestamp. If you want to filter out records that are older than timestamp 1589007600000, you could apply the following function:

Consider a table that has a string column campaign and a multi-value column double column prices. If you want to filter out records where campaign = 'X' or 'Y' and sum of all elements in prices is less than 100, you could apply the following function:

Filter config also supports SQL-like expression of inbuilt for filtering records (starting v 0.11.0+). Example:

Column Transformation

Transform functions can be defined on columns in the ingestion config of the table config.

For example, imagine that our source data contains the prices and timestamp fields. We want to extract the maximum price and store that in the maxPrices field and convert the timestamp into the number of hours since the epoch and store it in the hoursSinceEpoch field. You can do this by applying the following transformation:

Below are some examples of commonly used functions.

String concatenation

Concat firstName and lastName to get fullName

Find an element in an array

Find max value in array bids

Time transformation

Convert timestamp from MILLISECONDS to HOURS

Column name change

Change name of the column from user_id to userId

Extract value from a column containing space

Pinot doesn't support columns that have spaces, so if a source data column has a space, we'll need to store that value in a column with a supported name. To extract the value from first Name into the column firstName, run the following:

Ternary operation

If eventType is IMPRESSION set impression to 1. Similar for CLICK.

AVRO Map

Store an AVRO Map in Pinot as two multi-value columns. Sort the keys, to maintain the mapping. 1) The keys of the map as map_keys 2) The values of the map as map_values

Chaining transformations

Transformations can be chained. This means that you can use a field created by a transformation in another transformation function.

For example, we might have the following JSON document in the data field of our source data:

We can apply one transformation to extract the userId and then another one to pull out the numerical part of the identifier:

Flattening

There are 2 kinds of flattening:

One record into many

This is not natively supported as of yet. You can write a custom Decoder/RecordReader if you want to use this. Once the Decoder generates the multiple GenericRows from the provided input record, a List<GenericRow> should be set into the destination GenericRow, with the key $MULTIPLE_RECORDS_KEY$. The segment generation drivers will treat this as a special case and handle the multiple records case.

Extract attributes from complex objects

Feature TBD

Start Pinot components (scripts or docker images)

Setup Pinot by starting each component individually

Start Pinot Using Config Files

Often times we need to customized the setup of Pinot Components. Hence user can compile a config file and use it to start Pinot Components.

Below are the examples config files and sample command to start Pinot.

Pinot Controller

Below is a sample pinot-controller.conf used in HelmChart setup.

In order to run Pinot Controller, the command is:

Configure Controller

Below are some configurations you can set in Pinot Controller. You can head over to for complete list of available configs.

Pinot Broker

Below is a sample pinot-broker.conf used in HelmChart setup.

In order to run Pinot Broker, the command is:

Configure Broker

Below are some configurations you can set in Pinot Broker. You can head over to for complete list of available configs.

Pinot Server

Below is a sample pinot-server.conf used in HelmChart setup.

In order to run Pinot Server, the command is:

Configure Server

Below are some outstanding configurations you can set in Pinot Server. You can head over to for complete list of available configs.

Create and Configure table

A TABLE in regular database world is represented as <TABLE>_OFFLINE and/or <TABLE>_REALTIME in Pinot depending on the ingestion mode (batch, real-time, hybrid)

See for all possible batch/streaming tables.

Batch Table Creation

Please see for table configuration details and how to customize it.

Streaming Table Creation

Please see for table configuration details and how to customize it.

Load Data

Now that the table is configured, let's load some data. Data can be loaded in batch mode or streaming mode. See page for details. Loading data involves generating pinot segments from raw data and pushing them to the pinot cluster.

Load Data in Batch

User can always generate and push segments to Pinot via standalone scripts or using frameworks such as Hadoop or Spark. See this for more details on setting up Data Ingestion Jobs.

Below example goes with the standalone mode.

JobSpec yaml file has all the information regarding data format, input data location and pinot cluster coordinates. Note that this assumes that the controller is RUNNING to fetch the table config and schema. If not, you will have to configure the spec to point at their location. See for more details.

Load Data in Streaming

Kafka