Apache Pinot vs Apache Druid

Compare and contrast Apache Pinot and Apache Druid by architecture, ingestion, queries, performance, and scalability.

Compare Apache Pinot to Rockset here

Compare Apache Druid to Rockset here

Apache Pinot vs Apache Druid Architecture

Architecture

Apache Pinot

Apache Druid

Deployment model

PaaS or self managed

• SaaS or self managed. • Imply’s Druid cloud offering requires customers to configure, scale, and capacity plan

Use of storage hierarchy

Hot storage plus Deep Store for backup and restore operations

• Queries are served from data on disk and an in-memory cache • Cloud storage or HDFS for deep storage

Isolation of ingest and query

• Ingestion and queries are run on the same node by default • The ingestion and querying processes can be run on separate nodes, however not for real-time data

Separation of compute and storage

No, although StarTree separates compute and cloud storage, which is an important consideration for those considering StarTree vs. open-source Pinot

No, although Imply’s offering separates compute and storage

Isolation for multiple applications

Full isolation with replication

Pinot is a real-time distributed OLAP datastore that ingests both batch and streaming data. It has a distributed systems architecture that scales both horizontally and vertically, but unlike alternative OLAP databases, it does not decouple storage and compute. It supports both self-managed and PaaS options.

Druid’s architecture employs nodes called data servers that are used for both ingestion and queries. High ingestion or query load can cause CPU and memory contention compared with Druid alternatives. Breaking apart the pre-packaged ingestion and query server components involves planning ahead and additional complexity, and is not dynamic.

Apache Pinot vs Apache Druid Ingestion

Ingestion

Apache Pinot

Apache Druid

Data sources

Streaming • AWS Kinesis • Apache Kafka Batch • Cloud Storage • PostgreSQL • MySQL • Snowflake • Google BigQuery File upload Write API

Data connectors to: • Events streams (e.g. Kafka, Kinesis) • Data lakes (e.g. S3, Google Cloud Storage) • RDBMS and HDFS databases CDC events from databases require manual conversion to Druid events

Semi structured data

Support for JSON column types, but JSON must first be converted to a string

Druid requires flattening nested data at ingest and maintaining a flattening spec as the schema changes over time

Transformations and rollups

Yes

Yes, using ingestion specs written in JSON support rollups and simple transformations. SQL ingest transformations available for Imply

Pinot supports high-performance ingest from streaming data sources. Each table is either offline or real time. Real-time tables have a smaller retention period and scale based on ingestion rate while offline tables have a larger retention period and scale based on the amount of data. In order to persistently store the generated segments that make up a table, you will need to change controller and server configs to add deep storage.

Druid has built-in connectors that manage ingestion from common data sources. Unlike some Druid competitors, it doesn’t support nested data, so data must be flattened at ingest. Denormalization is also required at ingest, increasing operational burden for certain use cases.

Apache Pinot vs Apache Druid Performance

Performance

Apache Pinot

Apache Druid

Updates

• By default, all data in Pinot is immutable • Upserts only supported for streaming ingest • No support for upserts on data using the star tree index • No support for upserts on out-of-order events

Druid segments become immutable once committed and published, making it more appropriate for append-only use cases

Indexing

• Manually configured • Forward index, inverted Index, Star-Tree index (columnar), bloom filter, range index, search index, JSON index, geospatial index, timestamp index

• Bitmap index

Query latency

50-1000ms queries on 100s of TB

Typically sub-second query latency for denormalized, flattened datasets up to 100s of TBs

Storage format

Pinot stores data in a columnar format and adds additional indexes to perform fast filtering, aggregation and group by

Columnar format partitioned by time

Streaming ingest

1-2 second ingest for streaming data

• Sub 15 seconds

Like its competitors, Pinot can achieve sub-second query latency at high concurrency. However, this level of performance requires tuning, management, and deep expertise. Compared with the open-source version, the PaaS versions of Pinot address some of these issues, but similarly require expertise while making tradeoffs affecting query performance.

Druid is designed to make streaming data queryable as quickly as possible. JOINs are either impossible or incur a large performance penalty. Updates are only possible via batch jobs. Druid leverages data denormalization and write-time aggregation at ingestion to reduce query latency.

Apache Pinot vs Apache Druid Queries

Queries

Apache Pinot

Apache Druid

Joins

SQL joins limited to inner join, left-outer, and semi-join

Yes, only for datasets that fit in memory, and with a query latency penalty

Query language

SQL

Druid native queries • Druid SQL

Developer tooling

• API for querying SQL via POST command • Clients for JDBC, Java, Python, and Golang • Integrations with Trino and Presto

• Druid SQL API

Visualization tools

Integrations with Tableau and Superset

Pivot, maintained by Imply

In Pinot, SQL-like queries are received by brokers and scatter the request between real-time and offline servers. The two tables then process requests, send results back to the broker, and responds with the result. Joins are limited, as is support for UDFs and subqueries, making Pinot more or less useful depending on the use case.

Druid has a native JSON-based query language and provides Druid SQL as an alternative that translates into its native queries. JOINs are not recommended.

Apache Pinot vs Apache Druid Scalability

Scalability

Apache Pinot

Apache Druid

Vertical scaling

All 4 pinot node types can be resized manually

Users can manually increase or decrease server sizes

Horizontal scaling

• Users can add additional Pinot nodes to scale horizontally • Rebalancing is manual

• Users can manually add additional nodes to a scaled-out cluster. • Imply automates some of the horizontal scaling process.

Pinot allows for vertical scaling by increasing CPU and memory for each node as well as horizontal scaling by adding additional nodes. Capacity planning is a time-consuming, iterative, and manual task. It involves load testing and tuning across multiple vectors including read QPS, write QPS, number of streaming partitions, daily data size, retention period, types of workloads, number and type of segments, and much more.

Druid users are exposed to complex decisions about the number and size of servers as clusters are scaled.