Lookups

Lookups are a concept in Apache Druid where dimension values are (optionally) replaced with new values, allowing join-like functionality. Applying lookups in Druid is similar to joining a dimension table in a data warehouse. See dimension specs for more information. For the purpose of these documents, a "key" refers to a dimension value to match, and a "value" refers to its replacement. So if you wanted to map appid-12345 to Super Mega Awesome App then the key would be appid-12345 and the value would be Super Mega Awesome App.

It is worth noting that lookups support not just use cases where keys map one-to-one to unique values, such as country code and country name, but also support use cases where multiple IDs map to the same value, e.g. multiple app-ids mapping to a single account manager. When lookups are one-to-one, Druid is able to apply additional query rewrites; see below for more details.

Lookups do not have history. They always use the current data. This means that if the chief account manager for a particular app-id changes, and you issue a query with a lookup to store the app-id to account manager relationship, it will return the current account manager for that app-id REGARDLESS of the time range over which you query.

If you require data time range sensitive lookups, such a use case is not currently supported dynamically at query time, and such data belongs in the raw denormalized data for use in Druid.

Lookups are generally preloaded in-memory on all servers. But very small lookups (on the order of a few dozen to a few hundred entries) can also be passed inline in native queries time using the "map" lookup type. Refer to the dimension specs documentation for details.

Other lookup types are available as extensions, including:

Globally cached lookups from local files, remote URIs, or JDBC through lookups-cached-global.
Globally cached lookups from a Kafka topic through kafka-extraction-namespace.

info

Multi-value dimensions (MVDs) are not supported as keys in lookups. For example, to map the MVD ["A", "B", "C"] to the value x in your lookup, flatten the MVD and map each element of the MVD to the value. Your lookup will have separate key-value pairs for each element of the MVD: "A": "x", "B": "x", and "C": "x".

Query Syntax

In Druid SQL, lookups can be queried using the LOOKUP function, for example:

SELECT
  LOOKUP(store, 'store_to_country') AS country,
  SUM(revenue)
FROM sales
GROUP BY 1

The LOOKUP function also accepts a third argument called replaceMissingValueWith as a constant string. If the lookup does not contain a value for the provided key, then the LOOKUP function returns this replaceMissingValueWith value rather than NULL, just like COALESCE. For example, LOOKUP(store, 'store_to_country', 'NA') is equivalent to COALESCE(LOOKUP(store, 'store_to_country'), 'NA').

Lookups can be queried using the JOIN operator:

SELECT
  store_to_country.v AS country,
  SUM(sales.revenue) AS country_revenue
FROM
  sales
  INNER JOIN lookup.store_to_country ON sales.store = store_to_country.k
GROUP BY 1

info

The LOOKUP function has automatic query rewrites available that the JOIN approach does not, including reverse lookups and pulling up through GROUP BY. If these rewrites are important for you, consider using the LOOKUP function instead of JOIN.

In native queries, lookups can be queried with dimension specs or extraction functions.

Query Rewrites

Druid can perform two automatic query rewrites when using the LOOKUP function: reverse lookups and pulling up through GROUP BY. These rewrites and their requirements are described in the following sections.

Reverse lookup

When LOOKUP function calls appear in the WHERE clause of a query, Druid reverses them when possible. For example, if the lookup table sku_to_name contains the mapping 'WB00013' => 'WhizBang Sprocket', then Druid automatically rewrites this query:

SELECT
  LOOKUP(sku, 'sku_to_name') AS name,
  SUM(revenue)
FROM sales
WHERE LOOKUP(sku, 'sku_to_name') = 'WhizBang Sprocket'
GROUP BY LOOKUP(sku, 'sku_to_name')

Into this:

SELECT
  LOOKUP(sku, 'sku_to_name') AS name,
  SUM(revenue)
FROM sales
WHERE sku = 'WB00013'
GROUP BY LOOKUP(sku, 'sku_to_name')

The difference is that in the latter case, data servers do not need to apply the LOOKUP function while filtering, and can make more efficient use of indexes for sku.

The following table contains examples of when it is possible to reverse calls to `LOOKUP` while in Druid's default null handling mode. The list of examples is illustrative, albeit not exhaustive.

SQL	Reversible?
`LOOKUP(sku, 'sku_to_name') = 'WhizBang Sprocket'`	Yes
`LOOKUP(sku, 'sku_to_name') IS NOT DISTINCT FROM 'WhizBang Sprocket'`	Yes, for non-null literals
`LOOKUP(sku, 'sku_to_name') <> 'WhizBang Sprocket'`	No, unless `sku_to_name` is injective
`LOOKUP(sku, 'sku_to_name') IS DISTINCT FROM 'WhizBang Sprocket'`	Yes, for non-null literals
`LOOKUP(sku, 'sku_to_name') = 'WhizBang Sprocket' IS NOT TRUE`	Yes
`LOOKUP(sku, 'sku_to_name') IN ('WhizBang Sprocket', 'WhizBang Chain')`	Yes
`LOOKUP(sku, 'sku_to_name') NOT IN ('WhizBang Sprocket', 'WhizBang Chain')`	No, unless `sku_to_name` is injective
`LOOKUP(sku, 'sku_to_name') IN ('WhizBang Sprocket', 'WhizBang Chain') IS NOT TRUE`	Yes
`LOOKUP(sku, 'sku_to_name') IS NULL`	No
`LOOKUP(sku, 'sku_to_name') IS NOT NULL`	No
`LOOKUP(UPPER(sku), 'sku_to_name') = 'WhizBang Sprocket'`	Yes, to `UPPER(sku) = [key for 'WhizBang Sprocket']` (the `UPPER` function remains)
`COALESCE(LOOKUP(sku, 'sku_to_name'), 'N/A') = 'WhizBang Sprocket'`	Yes, but see next item for `= 'N/A'`
`COALESCE(LOOKUP(sku, 'sku_to_name'), 'N/A') = 'N/A'`	No, unless `sku_to_name` is injective, which allows Druid to ignore the `COALESCE`
`COALESCE(LOOKUP(sku, 'sku_to_name'), 'N/A') = 'WhizBang Sprocket' IS NOT TRUE`	Yes
`COALESCE(LOOKUP(sku, 'sku_to_name'), 'N/A') <> 'WhizBang Sprocket'`	Yes, but see next item for `<> 'N/A'`
`COALESCE(LOOKUP(sku, 'sku_to_name'), 'N/A') <> 'N/A'`	No, unless `sku_to_name` is injective, which allows Druid to ignore the `COALESCE`
`COALESCE(LOOKUP(sku, 'sku_to_name'), sku) = 'WhizBang Sprocket'`	No, `COALESCE` is only reversible when the second argument is a constant
`LOWER(LOOKUP(sku, 'sku_to_name')) = 'whizbang sprocket'`	No, functions other than `COALESCE` are not reversible
`MV_CONTAINS(LOOKUP(sku, 'sku_to_name'), 'WhizBang Sprocket')`	Yes
`NOT MV_CONTAINS(LOOKUP(sku, 'sku_to_name'), 'WhizBang Sprocket')`	No, unless `sku_to_name` is injective
`MV_OVERLAP(LOOKUP(sku, 'sku_to_name'), ARRAY['WhizBang Sprocket'])`	Yes
`NOT MV_OVERLAP(LOOKUP(sku, 'sku_to_name'), ARRAY['WhizBang Sprocket'])`	No, unless `sku_to_name` is injective

You can see the difference in the native query that is generated during SQL planning, which you can retrieve with EXPLAIN PLAN FOR. When a lookup is reversed in this way, the lookup function disappears and is replaced by a simpler filter, typically of type equals or in.

Lookups are not reversed if the number of matching keys exceeds the sqlReverseLookupThreshold or inSubQueryThreshold for the query.

This rewrite adds some planning time that may become noticeable for larger lookups, especially if many keys map to the same value. You can see the impact on planning time in the sqlQuery/planningTimeMs metric. You can also measure the time taken by EXPLAIN PLAN FOR, which plans the query but does not execute it.

This rewrite can be disabled by setting sqlReverseLookup: false in your query context.

Pull up

Lookups marked as injective can be pulled up through a GROUP BY. For example, if the lookup sku_to_name is injective, Druid automatically rewrites this query:

SELECT
  LOOKUP(sku, 'sku_to_name') AS name,
  SUM(revenue)
FROM sales
GROUP BY LOOKUP(sku, 'sku_to_name')

Into this:

SELECT
  LOOKUP(sku, 'sku_to_name') AS name,
  SUM(revenue)
FROM sales
GROUP BY sku

The difference is that the LOOKUP function is not applied until after the GROUP BY is finished, which speeds up the GROUP BY.

You can see the difference in the native query that is generated during SQL planning, which you can retrieve with EXPLAIN PLAN FOR. When a lookup is pulled up in this way, the lookup function call typically moves from the virtualColumns or dimensions section of a native query into the postAggregations.

This rewrite can be disabled by setting sqlPullUpLookup: false in your query context.

Injective lookups

Injective lookups are eligible for the largest set of query rewrites. Injective lookups must satisfy the following "one-to-one lookup" properties:

All values in the lookup table must be unique. That is, no two keys can map to the same value.
The lookup table must have a key-value pair defined for every input that the LOOKUP function call may encounter. For example, when calling LOOKUP(sku, 'sku_to_name'), the sku_to_name lookup table must have a key for all possible sku.
In SQL-compatible null handling mode (when druid.generic.useDefaultValueForNull = false, the default) injective lookup tables are not required to have keys for null, since LOOKUP of null is always null itself.
When druid.generic.useDefaultValueForNull = true, a LOOKUP of null retrieves the value mapped to the empty-string key (""). In this mode, injective lookup tables must have an empty-string key if the LOOKUP function may encounter null input values.

To determine whether a lookup is injective, Druid relies on an injective property that you can set in the lookup definition. In general, you should set injective: true for any lookup that satisfies the required properties, to allow Druid to run your queries as fast as possible.

Druid does not verify whether lookups satisfy these required properties. Druid may return incorrect query results if you set injective: true for a lookup table that is not actually a one-to-one lookup.

Dynamic Configuration

The following documents the behavior of the cluster-wide config which is accessible through the Coordinator. The configuration is propagated through the concept of "tier" of servers. A "tier" is defined as a group of services which should receive a set of lookups. For example, you might have all Historicals be part of __default, and Peons be part of individual tiers for the datasources they are tasked with. The tiers for lookups are completely independent of Historical tiers.

These configs are accessed using JSON through the following URI template

http://<COORDINATOR_IP>:<PORT>/druid/coordinator/v1/lookups/config/{tier}/{id}

All URIs below are assumed to have http://<COORDINATOR_IP>:<PORT> prepended.

If you have NEVER configured lookups before, you MUST post an empty json object {} to /druid/coordinator/v1/lookups/config to initialize the configuration.

These endpoints will return one of the following results:

404 if the resource is not found
400 if there is a problem in the formatting of the request
202 if the request was accepted asynchronously (POST and DELETE)
200 if the request succeeded (GET only)

Configuration propagation behavior

The configuration is propagated to the query serving processes (Broker / Router / Peon / Historical) by the Coordinator. The query serving processes have an internal API for managing lookups on the process and those are used by the Coordinator. The Coordinator periodically checks if any of the processes need to load/drop lookups and updates them appropriately.

Please note that only 2 simultaneous lookup configuration propagation requests can be concurrently handled by a single query serving process. This limit is applied to prevent lookup handling from consuming too many server HTTP connections.

API

See Lookups API for reference on configuring lookups and lookup status.

Configuration

See Lookups Dynamic Configuration for Coordinator configuration.

To configure a Broker / Router / Historical / Peon to announce itself as part of a lookup tier, use following properties.

Property	Description	Default
`druid.lookup.lookupTier`	The tier for lookups for this process. This is independent of other tiers.	`__default`
`druid.lookup.lookupTierIsDatasource`	For some things like indexing service tasks, the datasource is passed in the runtime properties of a task. This option fetches the tierName from the same value as the datasource for the task. It is suggested to only use this as Peon options for the indexing service, if at all. If true, `druid.lookup.lookupTier` MUST NOT be specified	`"false"`

To configure the behavior of the dynamic configuration manager, use the following properties on the Coordinator:

Property	Description	Default
`druid.manager.lookups.hostTimeout`	Timeout (in ms) PER HOST for processing request	`2000`(2 seconds)
`druid.manager.lookups.allHostTimeout`	Timeout (in ms) to finish lookup management on all the processes.	`900000`(15 mins)
`druid.manager.lookups.period`	How long to pause between management cycles	`120000`(2 mins)
`druid.manager.lookups.threadPoolSize`	Number of service processes that can be managed concurrently	`10`

Saving configuration across restarts

It is possible to save the configuration across restarts such that a process will not have to wait for Coordinator action to re-populate its lookups. To do this the following property is set:

Property	Description	Default
`druid.lookup.snapshotWorkingDir`	Working path used to store snapshot of current lookup configuration, leaving this property null will disable snapshot/bootstrap utility	null
`druid.lookup.enableLookupSyncOnStartup`	Enable the lookup synchronization process with Coordinator on startup. The queryable processes will fetch and load the lookups from the Coordinator instead of waiting for the Coordinator to load the lookups for them. Users may opt to disable this option if there are no lookups configured in the cluster.	true
`druid.lookup.numLookupLoadingThreads`	Number of threads for loading the lookups in parallel on startup. This thread pool is destroyed once startup is done. It is not kept during the lifetime of the JVM	Available Processors / 2
`druid.lookup.coordinatorFetchRetries`	How many times to retry to fetch the lookup bean list from Coordinator, during the sync on startup.	3
`druid.lookup.lookupStartRetries`	How many times to retry to start each lookup, either during the sync on startup, or during the runtime.	3
`druid.lookup.coordinatorRetryDelay`	How long to delay (in millis) between retries to fetch lookup list from the Coordinator during the sync on startup.	60_000

Introspect a Lookup

The Broker provides an API for lookup introspection if the lookup type implements a LookupIntrospectHandler.

A GET request to /druid/v1/lookups/introspect/{lookupId} will return the map of complete values.

ex: GET /druid/v1/lookups/introspect/nato-phonetic

{
    "A": "Alfa",
    "B": "Bravo",
    "C": "Charlie",
    ...
    "Y": "Yankee",
    "Z": "Zulu",
    "-": "Dash"
}

The list of keys can be retrieved via GET to /druid/v1/lookups/introspect/{lookupId}/keys"

ex: GET /druid/v1/lookups/introspect/nato-phonetic/keys

[
    "A",
    "B",
    "C",
    ...
    "Y",
    "Z",
    "-"
]

A GET request to /druid/v1/lookups/introspect/{lookupId}/values" will return the list of values.

ex: GET /druid/v1/lookups/introspect/nato-phonetic/values

[
    "Alfa",
    "Bravo",
    "Charlie",
    ...
    "Yankee",
    "Zulu",
    "Dash"
]

Query Syntax​

Query Rewrites​

Reverse lookup​

Pull up​

Injective lookups​

Dynamic Configuration​

Configuration propagation behavior​

API​

Configuration​

Saving configuration across restarts​

Introspect a Lookup​