Pinot is designed to deliver low latency queries on large datasets. In order to achieve this performance, Pinot stores data in a columnar format and adds additional indices to perform fast filtering, aggregation and group by.

Raw data is broken into small data shards and each shard is converted into a unit known as a segment. One or more segments together form a table, which is the logical container for querying Pinot using SQL/PQL.

Pinot Storage Model

Pinot uses a variety of terms that can refer to either abstractions that model the storage of data or infrastructure components that drive the functionality of the system.

Table

Similar to traditional databases, Pinot has the concept of a table—a logical abstraction to refer to a collection of related data.

As is the case with RDBMS, a table is a construct that consists of columns and rows (documents) that are queried using SQL. A table is associated with a schema that defines the columns in a table as well as their data types.

In contrast to RDBMS schemas, multiple tables in Pinot (real-time or batch) can inherit a single schema definition. Tables are independently configured for concerns such as indexing strategies, partitioning, tenants, data sources, and/or replication.

Segment

Pinot has a distributed systems architecture that scales horizontally. Pinot expects the size of a table to grow infinitely over time. In order to achieve this, all data needs to be distributed across multiple nodes. Pinot achieves this by breaking data into smaller chunks known as segments (similar to shards/partitions in HA relational databases). Segments can also be seen as time-based partitions.

Tenant

In order to support multi-tenancy, Pinot has first class support for tenants. A table is associated with a tenant. This allows all tables belonging to a particular logical namespace to be grouped under a single tenant name and isolated from other tenants. This isolation between tenants provides different namespaces for applications and teams to prevent sharing tables or schemas. Development teams building applications will never have to operate an independent deployment of Pinot. An organization can operate a single cluster and scale it out as new tenants increase the overall volume of queries. Developers can manage their own schemas and tables without being impacted by any other tenant on a cluster.

By default, all tables belong to a default tenant named "default". The concept of tenants is very important, as it satisfies the architectural principle of a "database per service/application" without having to operate many independent data stores. Further, tenants will schedule resources so that segments (shards) are able to restrict a table's data to reside only on a specified set of nodes. Similar to the kind of isolation that is ubiquitously used in Linux containers, compute resources in Pinot can be scheduled to prevent resource contention between tenants.

Cluster

Logically, a cluster is simply a group of tenants. As with the classical definition of a cluster, it is also a grouping of a set of compute nodes. Typically, there is only one cluster per environment/data center. There is no needed to create multiple clusters since Pinot supports the concept of tenants. At LinkedIn, the largest Pinot cluster consists of 1000+ nodes distributed across a data center. The number of nodes in a cluster can be added in a way that will linearly increase performance and availability of queries. The number of nodes and the compute resources per node will reliably predict the QPS for a Pinot cluster, and as such, capacity planning can be easily achieved using SLAs that assert performance expectations for end-user applications.

Pinot Components

A Pinot cluster comprises multiple distributed system components. These components are useful to understand for operators that are monitoring system usage or are debugging an issue with a cluster deployment.

• Controller

• Server

• Broker

• Minion (optional)

The benefits of scale that make Pinot linearly scalable for an unbounded number of nodes is made possible through its integration with Apache Zookeeper and Apache Helix.

Pinot Controller

A controller is the core orchestrator that drives the consistency and routing in a Pinot cluster. Controllers are horizontally scaled as an independent component (container) and has visibility of the state of all other components in a cluster. The controller reacts and responds to state changes in the system and schedules the allocation of resources for tables, segments, or nodes. As mentioned earlier, Helix is embedded within the controller as an agent that is a participant responsible for observing and driving state changes that are subscribed to by other components.

In addition to cluster management, resource allocation, and scheduling, the controller is also the HTTP gateway for REST API administration of a Pinot deployment. A web-based query console is also provided for operators to quickly and easily run SQL/PQL queries.

Pinot Broker

A broker receives queries from a client and routes their execution to one or more Pinot servers before returning a consolidated response.

Pinot Server

Servers host segments (shards) that are scheduled and allocated across multiple nodes and routed on an assignment to a tenant (there is a single-tenant by default). Servers are independent containers that scale horizontally and are notified by Helix through state changes driven by the controller. A server can either be a real-time server or an offline server.

A real-time and offline server have very different resource usage requirements, where real-time servers are continually consuming new messages from external systems (such as Kafka topics) that are ingested and allocated on segments of a tenant. Because of this, resource isolation can be used to prioritize high-throughput real-time data streams that are ingested and then made available for query through a broker.

Pinot Minion

Pinot minion is an optional component that can be used to run background tasks such as "purge" for GDPR (General Data Protection Regulation). As Pinot is an immutable aggregate store, records containing sensitive private data need to be purged on a request-by-request basis. Minion provides a solution for this purpose that complies with GDPR while optimizing Pinot segments and building additional indices that guarantee performance in the presence of the possibility of data deletion. One can also write a custom task that runs on a periodic basis. While it's possible to perform these tasks on the Pinot servers directly, having a separate process (Minion) lessens the overall degradation of query latency as segments are impacted by mutable writes.

Join us in our Slack channel for questions, troubleshooting, and feedback. You can request an invite from - https://communityinviter.com/apps/apache-pinot/apache-pinot.

Pinot is a real-time distributed OLAP datastore, purpose-built to provide ultra low-latency analytics, even at extremely high throughput. It can ingest directly from streaming data sources - such as Apache Kafka and Amazon Kinesis - and make the events available for querying instantly. It can also ingest from batch data sources such as Hadoop HDFS, Amazon S3, Azure ADLS, and Google Cloud Storage.

At the heart of the system is a columnar store, with several smart indexing and pre-aggregation techniques for low latency. This makes Pinot the most perfect fit for user-facing realtime analytics. At the same time, Pinot is also a great choice for other analytical use-cases, such as internal dashboards, anomaly detection, and ad-hoc data exploration.

Pinot was built by engineers at LinkedIn and Uber and is designed to scale up and out with no upper bound. Performance always remains constant based on the size of your cluster and an expected query per second (QPS) threshold.

User-Facing Real-Time Analytics

User-facing analytics, or site-facing analytics, is the analytical tools and applications that you would expose directly to the end-users of your product. In a user-facing analytics application, think of the user-base as ALL end users of an App. This App could be a social networking app, or a food delivery app - anything at all. It’s not just a few analysts doing offline analysis, or a handful of data scientists in a company running ad-hoc queries. This is ALL end-users, receiving personalized analytics on their personal devices (think 100s of 1000s of queries per second). These queries are triggered by apps, and not written by people, and so the scale will be as much as the active users on that App (think millions of events/sec)

And, this is for all the freshest possible data, which touches on the other aspect here - realtime analytics. "Yesterday" might be a long time ago for some businesses and they cannot wait for ETLs and batch jobs. The data needs to be used for analytics, as soon as it is generated (think latencies < 1s).

Why is user-facing real-time analytics is so challenging?

Wanting such a user-facing analytics application, using realtime events, sounds great. But what does it mean for the underlying infrastructure, to support such an analytical workload?

1. Such applications require the freshest possible data, and so the system needs to be able to ingest data in real time and make it available for querying, also in real time.

2. Data for such apps tend to be event data, for a wide range of actions, coming from multiple sources, and so the data comes in at a very high velocity and tends to be highly dimensional.

3. Queries are triggered by end-users interacting with apps - with queries per second in hundreds of thousands, with arbitrary query patterns, and latencies are expected to be in milliseconds for good user-experience.

4. And further do all of the above, while being scalable, reliable, highly available, and having a low cost to serve.

This video talks more about user-facing real-time analytics, and how Pinot is used to achieve that.

Here's another great video that goes into the details of how Pinot tackles some of the challenges faced in handling a user-facing analytics workload.

Companies using Pinot

Pinot originated at LinkedIn which currently has one of the largest deployment powering more than 50+ user facing applications such as Viewed My Profile, Talent Analytics, Company Analytics, Ad Analytics and many more. At LinkedIn, Pinot also serves as the backend to visualize and monitor 10,000+ business metrics.

With Pinot's growing popularity, several companies are now using it in production to power a variety of analytics use cases. A detailed list of companies using Pinot can be found here.​

Features

• A column-oriented database with various compression schemes such as Run Length, Fixed Bit Length

• Pluggable indexing technologies - Sorted Index, Bitmap Index, Inverted Index, StarTree Index, Bloom Filter, Range Index, Text Search Index(Lucence/FST), Json Index, Geospatial Index

• Ability to optimize query/execution plan based on query and segment metadata

• Near real-time ingestion from streams such as Kafka, Kinesis and batch ingestion from sources such as Hadoop, S3, Azure, GCS

• SQL-like language that supports selection, aggregation, filtering, group by, order by, distinct queries on data

• Support for multi-valued fields

• Horizontally scalable and fault-tolerant

When should I use it?

Pinot is designed to execute OLAP queries with low latency. It is suited in contexts where fast analytics, such as aggregations, are needed on immutable data, possibly, with real-time data ingestion.

User facing Analytics Products

Pinot is the perfect choice for user-facing analytics products. Pinot was originally built at LinkedIn to power rich interactive real-time analytic applications such as Who Viewed Profile, Company Analytics, Talent Insights, and many more. UberEats Restaurant Manager is another example of a customer-facing Analytics App. At LinkedIn, Pinot powers 50+ user-facing products, ingesting millions of events per second and serving 100k+ queries per second at millisecond latency.

Real-time Dashboard for Business Metrics

Pinot can be also be used to perform typical analytical operations such as slice and dice, drill down, roll up, and pivot on large scale multi-dimensional data. For instance, at LinkedIn, Pinot powers dashboards for thousands of business metrics. One can connect various BI tools such as Superset, Tableau, or PowerBI to visualize data in Pinot.

Instructions to connect Pinot with Superset can be found here.

Anomaly Detection

In addition to visualizing data in Pinot, one can run Machine Learning Algorithms to detect Anomalies in the data stored in Pinot. See ThirdEye for more information on how to use Pinot for Anomaly Detection and Root Cause Analysis.

Frequently asked questions when getting started

Is Pinot a data warehouse or a database?

While Pinot doesn't match the typical mold of a database product, it is best understood based on your role as either an analyst, data scientist, or application developer.

Enterprise business intelligence

For analysts and data scientists, Pinot is best viewed as a highly-scalable data platform for business intelligence. In this view, Pinot converges big data platforms with the traditional role of a data warehouse, making it a suitable replacement for analysis and reporting.

Enterprise application development

For application developers, Pinot is best viewed as an immutable aggregate store that sources events from streaming data sources, such as Kafka, and makes it available for a query using SQL.

As is the case with a microservice architecture, data encapsulation ends up requiring each application to provide its own data store, as opposed to sharing one OLTP database for reads and writes. In this case, it becomes difficult to query the complete view of a domain because it becomes stored in many different databases. This is costly in terms of performance since it requires joins across multiple microservices that expose their data over HTTP under a REST API. To prevent this, Pinot can be used to aggregate all of the data across a microservice architecture into one easily queryable view of the domain.

Pinot tenants prevent any possibility of sharing ownership of database tables across microservice teams. Developers can create their own query models of data from multiple systems of record depending on their use case and needs. As with all aggregate stores, query models are eventually consistent and immutable.

Get started

Our documentation is structured to let you quickly get to the content you need and is organized around the different concerns of users, operators, and developers. If you're new to Pinot and want to learn things by example, please take a look at our getting started section.

Starter guides

To start importing data into Pinot, check out our guides on batch import and stream ingestion based on our plugin architecture.

Query example

Pinot works very well for querying time series data with many dimensions and metrics over a vast unbounded space of records that scales linearly on a per-node basis. Filters and aggregations are both easy and fast.

SELECT sum(clicks), sum(impressions) FROM AdAnalyticsTable
  WHERE 
       ((daysSinceEpoch >= 17849 AND daysSinceEpoch <= 17856)) AND 
       accountId IN (123456789)
  GROUP BY 
       daysSinceEpoch TOP 100

Pinot supports SQL for querying read-only data. Learn more about querying Pinot for time series data in our PQL (Pinot Query Language) guide.

Installation

Pinot may be deployed to and operated on a cloud provider or a local or virtual machine. You may get started either with a bare-metal installation or a Kubernetes one (either locally or in the cloud). To get immediately started with Pinot, check out these quick start guides for bootstrapping a Pinot cluster using Docker or Kubernetes.

Standalone mode

Cluster mode

Learn

For a high-level overview that explains how Pinot works, please take a look at our basic concepts section.

To understand the distributed systems architecture that explains Pinot's operating model, please take a look at our basic architecture section.

This page will introduce you to the guiding principles behind the design of Apache Pinot. Here you will learn the distributed systems architecture that allows Pinot to scale the performance of queries linearly based on the number of nodes in a cluster. You'll also be introduced to the two different types of tables used to ingest and query data in offline (batch) or real-time (stream) mode.

Guiding design principles

Pinot was designed by engineers at LinkedIn and Uber to scale query performance based on the number of nodes in a cluster. As you add more nodes, query performance will always improve based on the expected query volume per second quota. To achieve horizontal scalability to an unbounded number of nodes and data storage, without performance degradation, the following guiding design principles were established.

• Highly available: Pinot is built to serve low latency analytical queries for customer facing applications. By design, there is no single point of failure in Pinot. The system continues to serve queries when a node goes down.

• Horizontally scalable: Ability to scale by adding new nodes as a workload changes.

• Latency vs Storage: Pinot is built to provide low latency even at high-throughput. Features such as segment assignment strategy, routing strategy, star-tree indexing were developed to achieve this.

• Immutable data: Pinot assumes that all data stored is immutable. For GDPR compliance, we provide an add-on solution for purging data while maintaining performance guarantees.

• Dynamic configuration changes: Operations such as adding new tables, expanding a cluster, ingesting data, modifying indexing config, and re-balancing must be performed without impacting query availability or performance.

Core components

Apache Helix and Zookeeper

Helix divides nodes into three logical components based on their responsibilities:

1. Participant: These are the nodes in the cluster that actually host the distributed storage resources.

2. Spectator: These nodes observe the current state of each participant and routes requests accordingly. Routers, for example, need to know the instance on which a partition is hosted and its state in order to route the request to the appropriate endpoint. Routing is continually being changed to optimize cluster performance as storage primitives are added and changed.

Helix uses Zookeeper to maintain cluster state. Each component in a Pinot cluster takes a Zookeeper address as a startup parameter. The various components that are distributed in a Pinot cluster will watch Zookeeper notifications and issue updates via its embedded Helix-defined agent.

Helix agents use Zookeeper to store and update configurations, as well as for distributed coordination. Zookeeper stores the following information about the cluster:

Knowing the ZNode layout structure in Zookeeper for Helix agents in a cluster is useful for operations and/or troubleshooting cluster state and health.

Controller

Fault tolerance

To achieve fault tolerance, one can start multiple controllers (typically three) and one of them will act as a leader. If the leader crashes or dies, another leader is automatically elected. Leader election is achieved using Apache Helix. Having at-least one controller is required to perform any DDL equivalent operation on the cluster, such as adding a table or a segment.

The controller does not interfere with query execution. Query execution is not impacted even when all controllers nodes are offline. If all controller nodes are offline, the state of the cluster will stay as it was when the last leader went down. When a new leader comes online, a cluster resumes re-balancing activity and can accept new tables or segments.

Controller REST interface

Broker

Brokers need three key things to start.

• Cluster name

• Zookeeper address

• Broker instance name

At the start, a broker registers as a Helix Participant and awaits notifications from other Helix agents. These notifications will be handled for table creation, a new segment being loaded, or a server starting up/or going down, in addition to any configuration changes.

Service Discovery/Routing Table

Irrespective of the kind of notification, the key responsibility of a broker is to maintain the query routing table. The query routing table is simply a mapping between segments and the servers that a segment resides on. Typically, a segment resides on more than one server. The broker computes multiple routing tables depending on the configured routing strategy for a table. The default strategy is to balance the query load across all available servers.

Query processing

For every query, a cluster's broker performs the following:

• Scatter-Gather: sends the requests to each server and gathers the responses.

• Merge: merges the query results returned from each server.

• Sends the query result to the client.

Fault tolerance

Broker instances scale horizontally without an upper bound. In a majority of cases, only three brokers are required. If most query results that are returned to a client are <1MB in size per query, one can run a broker and servers inside the same instance container. This lowers the overall footprint of a cluster deployment for use cases that do not need to guarantee a strict SLA on query performance in production.

Server

Offline servers

Real-time servers

Minion

Data ingestion overview

The two types of tables also scale differently.

• Real-time tables have a smaller retention period and scales query performance based on the ingestion rate.

• Offline tables have larger retention and scales performance based on the size of stored data.

Batch data flow

Real-time data flow

At table creation, a controller creates a new entry in Zookeeper for the consuming segment. Helix notices the new segment and notifies the real-time server, which starts consuming data from the streaming source. The broker, which watches for changes, detects the new segments and adds them to the list of segments to query (segment-to-server routing table).

Whenever the segment is complete (i.e. full), the real-time server notifies the Controller, which checks with all replicas and picks a winner to commit the segment to. The winner commits the segment and uploads it to the cluster's segment store, updating the state of the segment from "consuming" to "online". The controller then prepares a new segment in a "consuming" state.

Query overview

Queries are received by brokers—which checks the request against the segment-to-server routing table—scattering the request between real-time and offline servers.

The two tables then process the request by filtering and aggregating the queried data, which is then returned back to the broker. Finally, the broker gathers together all of the pieces of the query response and responds back to the client with the result.

This section is a reference for the definition of major components and logical abstractions used in Pinot.

Operator reference

Developer reference