Introduction to projections

Projections is an EventStoreDB subsystem that lets you append new events or link existing events to streams in a reactive manner.

Projections are good at solving one specific query type, a category known as 'temporal correlation queries'. This query type is common in business systems and few can execute these queries well.


Projections require the event body to be in JSON.

Business case examples

For example. You are looking for how many Twitter users said "happy" within 5 minutes of the word "foo coffee shop" and within 2 minutes of saying "london".

This is the type of query that projections can solve. Let's try a more complex business problem.

As a medical research doctor you want to find people diagnosed with pancreatic cancer within the last year. During their treatment a patient should not have had any proxies for a heart condition such as taking aspirin every morning. Within three weeks of their diagnosis they should have been put on treatment X. Within one month after starting the treatment they should have failed with a lab result that looks like L1. Within another six weeks they should have been put on treatment Y, and within four weeks failed that treatment with a lab result that looks like L2.

You can use projections in nearly all examples of near real-time complex event processing. There are a large number of problems that fit into this category from monitoring of temperature sensors, to reacting to changes in the stock market.

It's important to remember the types of problems that projections help to solve. Many problems are not a good fit for projections and are better served by hosting another read model populated by a catchup subscription.

Continuous querying

Projections support the concept of continuous queries. When running a projection you can choose whether the query should run and give you all results present, or whether the query should continue running into the future finding new results as they happen and updating its result set.

In the medical example above the doctor could leave the query running to be notified of any new patients that meet the criteria. The output of all queries is a stream, you can listen to this stream like any other stream.

Types of projections

There are two types of projections in EventStoreDB:

Performance impact

Keep in mind that all projections emit events as a reaction to events that they process. We call this effect write amplification because emitting new events or link events creates additional load on the server IO.

Some system projections emit link events to their streams for each event appended to the database. These projections are By Category, By Event Type and By Correlation Id. If all those three projections are enabled and started, adding one event to the database will, in fact, produce three additional events and, therefore, quadruples the number of write operations.

System projections $streams and $stream-by-category produce new events too, either per each new stream or per new stream category. If your system has a lot of small streams, the $streams system projection would also amplify writes significantly.

Custom projections create the most significant write amplification since they produce new events or link events, which in turn get processed by system projections.

Projections only run on a leader node of the cluster due to consistency concerns. It creates more CPU and IO load on the leader node compared to follower nodes.


Streams where projections emit events cannot be used to append events from applications. When this happens, the projection will detect events not produced by the projection itself and it will break.

The reason projections exclusively own their streams is because otherwise they would lose all predictability. The projection would no longer have any idea what should be in that stream. For example, when a projection starts up from a checkpoint, it first goes through all the events after that checkpoint and checks them against the emitted stream. By doing this, the projection can understand if it up to the last event and can continue from where it left off. On top of that, the projection can verify that everything is in order, no events missing, etc. If anyone can append to the emitted streams, then the projection would have no idea where it got to last in terms of processing. Therefore, it can no longer trust that the projection itself emitted that event or if something else did.