The general explanation of the multi-stage query engine is provided in the reference documentation. This section provides a deep dive into the multi-stage query engine. Most of the concepts explained here are related to the internals of the multi-stage query engine and users don't need to know about them in order to write queries. However, understanding these concepts can help you to take advantage of the engine's capabilities and to troubleshoot issues.

The multi-stage query engine uses a set of operators to process the query. These operators are based on relational algebra, with some modifications to better fit the distributed nature of the engine.

These operators are the execution units that Pinot uses to execute a query. The operators are executed in a pipeline with tree structure, where each operator consumes the output of the previous operators (also known as upstreams).

Users do not directly specify these operators. Instead they write SQL queries that are translated into a logical plan, which is then transformed into different operators. The logical plan can be obtained using explaining a query, while there is no way to get the operators directly. The closest thing to the operators that users can get is the multi-stage stats output.

Operators vs SQL clauses

These operators are generated from the SQL query that you write, but even they are similar, there is not a one-to-one mapping between the SQL clauses and the operators. Some SQL clauses generate multiple operators, while some operators are generated by multiple SQL clauses.

Operators vs explain plan nodes

Operators and explain plan nodes are closer than SQL clauses and operators. Although most explain plan nodes can be directly mapped to an operator, there are some exceptions:

• Each PinotLogicalExchange and each PinotLogicalSortExchange explain node is materialized into a pair of and operators.

• All plan nodes that belong to the same leaf stage are executed in the operator.

In general terms, the operators are the execution units that Pinot uses to execute a query and are also known as the multi-stage physical plan, while the explain plan nodes are logical plans. The difference between the two is that the operators can be actually executed, while the explain plan nodes are the logical representation of the query plan.

List of operators

The following is a list of operators that are used by the multi-stage query engine:

The filter operator is used to filter rows based on a condition.

This page describes the filter operator defined in the relational algebra used by multi-stage queries. This operator is generated by the multi-stage query engine when you use the where, having or sometimes on clauses.

Implementation details

Filter operations apply a predicate to each row and only keep the rows that satisfy the predicate.

It is important to notice that filter operators can only be optimized using indexes when they are executed in the leaf stage. The reason for that is that the intermediate stages don't have access to the actual segments. This is why the engine will try to push down the filter operation to the leaf stage whenever possible.

As explained in , the explain plan in the multi-stage query engine does not indicate whether indexes are used or not.

Blocking nature

The filter operator is a streaming operator. It emits the blocks of rows as soon as they are received from the upstream operator.

Hints

None

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1. This number is affected by the number of received rows and the complexity of the predicate.

emittedRows

Type: Long

The number of groups emitted by the operator. A large number of emitted rows may not be problematic, but indicates that the predicate is not very selective.

Explain attributes

The filter operator is represented in the explain plan as a LogicalFilter explain node.

condition

Type: Expression

The condition that is being applied to the rows. The expression may use indexed columns ($0, $1, etc), functions and literals. The indexed columns are always 0-based.

For example, the following explain plan:

Is saying that the filter is applying the condition $5 > 2 which means that only the rows where the 6th column is greater than 2 will be emitted. In order to know which column is the 6th, you need to look at the schema of the table scanned.

Tips and tricks

How to know if indexes are used

As explained in , the explain plan in the multi-stage query engine does not directly indicate whether indexes are used or not.

Apache Pinot contributors are working on improving this, but it is not yet available. Meanwhile, we need an indirect approach to get that information.

First, we need to know on which stage the filter is being used. If the filter is being used in an intermediate stage, then the filter is not using indexes. In order to know the stage, you can extract stages as explained in .

But what about the leaf filters executed in the stage? Not all filters in the leaf stage can use indexes. The only way to know if the filter is using indexes is to use single-stage explain plan. In order to do so you need to transform the leaf stage into a single-stage query. This is a manual process that can be tedious but ends up not being so difficult once you get used to it.

See for more information.

The aggregate operator is used to perform calculations on a set of rows and return a single row of results.

This page describes the aggregate operator defined in the relational algebra used by multi-stage queries. This operator is generated by the multi-stage query engine when you use aggregate functions in a query either with or without a group by clause.

Implementation details

Aggregate operations may be expensive in terms of memory, CPU and network usage. As explained in , the multi-stage query engine breaks down the query into multiple stages and each stage is then executed in parallel on different workers. Each worker processes a subset of the data and sends the results to the coordinator which then aggregates the results. When possible, the multi-stage query engine will try to apply a divide-and-conquer strategy to reduce the amount of data that needs to be processed in the coordinator stage.

For example if the aggregation function is a sum, the engine will try to sum the results of each worker before sending the partial result to the coordinator, which would then sum the partial results in order to get the final result. But some aggregation functions, like count(distinct), cannot be computed in this way and require all the data to be processed in the coordinator stage.

In Apache Pinot 1.1.0, the multi-stage query engine always keeps the data in memory. This means that the amount of memory used by the engine is proportional to the number of groups generated by the group by clause and the amount of data that needs to be kept for each group (which depends on the aggregation function).

Even when the aggregation function is a simple count, which only requires to keep a long for each group in memory, the amount of memory used can be high if the number of groups is high. This is why the engine limits the number of groups. By default, this limit is 100.000, but this can be changed by providing hints.

Blocking nature

The aggregate operator is a blocking operator. It needs to consume all the input data before emitting the result.

Hints

num_groups_limit

Type: Integer

Default: 100.000

Defines the max number of groups that can be created by the group by clause. If the number of groups exceeds this limit, the query will not fail but will stop the execution.

Example:

is_partitioned_by_group_by_keys

Type: Boolean

Default: false

If set to true, the engine will consider that the data is already partitioned by the group by keys. This means that the engine will not need to shuffle the data to group them by the group by keys and the coordinator stage will be able to compute the final result without needing to merge the partial results.

Example:

is_skip_leaf_stage_group_by

Type: Boolean

Default: false

If set to true, the engine will not push down the aggregate into the leaf stage. In some situations, it could be wasted effort to do group-by on leaf, eg: when cardinality of group by column is very high.

Example:

max_initial_result_holder_capacity

Type: Integer

Default: 10.000

Defines the initial capacity of the result holder that stores the intermediate results of the aggregation. This hint can be used to reduce the memory usage of the engine by setting a value close to the expected number of groups. It is usually recommended to not change this hint unless you know that the expected number of groups is much lower than the default value.

Example:

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1. Remember that this value is affected by the number of received rows and the complexity of the aggregation function.

emittedRows

Type: Long

The number of groups emitted by the operator. A large number of emitted rows can indicate that the query is not well optimized.

Remember that the number of groups is limited by the num_groups_limit hint and a large number of groups can lead to high memory usage and slow queries.

numGroupsLimitReached

Type: Boolean

This stat is set to true when the number of groups exceeds the limit defined by the num_groups_limit hint. In that case, the query will not fail but will return partial results, which will be indicated by the global partialResponse stat.

Explain attributes

The aggregate operator is represented in the explain plan as a LogicalAggregate explain node.

Remember that these nodes appear in pairs: First in one stage where the aggregation is done in parallel and then in the upstream stage where the partial results are merged.

group

Type: List of Integer

The list of columns used in the group by clause. These numbers are 0-based column indexes on the virtual row projected by the upstream.

For example the explain plan:

Is saying that the group by clause is using the column with index 6 in the virtual row projected by the upstream. Given in this case that row is generated by a table scan, the column with index 6 is the 7th column in the table as defined in its schema.

agg#N

Type: Expression

The aggregation functions applied to the columns. There may be multiple agg#N attributes, each one representing a different aggregation function.

For example the explain plan:

Has two aggregation functions: COUNT() and MAX(). The second is applied to the column with index 5 in the virtual row projected by the upstream. Given in this case that row is generated by a table scan, the column with index 5 is the 6th column in the table as defined in its schema.

Tips and tricks

Try to not use aggregation functions that cannot be parallelized

For example, it is recommended to use one of the different hyperloglog flavor instead of count(distinct) when the cardinality of the data or their size.

For example, it is cheaper to execute count(distinct) on an int column with 1000 distinct values than on a column that stores very long strings, even if the number of distinct values is the same.

The literal operator is used to define a constant value in the query. This operator may be generated by the multi-stage query engine when you use a constant value in a query.

Blocking nature

The literal operator is a blocking operator, but given its trivial nature it should not matter.

Implementation details

The literal operator is a simple operator that does not require any computation.

Hints

None

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1.

emittedRows

Type: Long

The number of groups emitted by the operator. It should always be one.

Explain attributes

None

Tips and tricks

Avoid very large literals

Take care when using very large literals (in the order hundreds of KBs), as they may need to be sent from brokers to servers and in general may introduce latencies in the parsing and query optimization.

The mailbox send operator is the operator that sends data to the mailbox receive operator. This is not an actual relational operator but a Pinot extension used to send data to other stages.

These operators are always the root of the intermediate and leaf stages.

Implementation details

Stages in the multi-stage query engine are executed in parallel by different workers. Workers send data to each other using mailboxes. The number of mailboxes depends on the send operator parallelism, the receive operator parallelism and the distribution being used. At worse, there is one mailbox per worker pair, so if the upstream send operator has a parallelism of S and the receive operator has a parallelism of R, there will be S * R mailboxes.

The mailbox receive operator is the operator that receives the data from the mailbox send operator. This is not an actual relational operator but a Pinot extension used to receive data from other stages.

Implementation details

Stages in the multi-stage query engine are executed in parallel by different workers. Workers send data to each other using mailboxes. The number of mailboxes depends on the send operator parallelism, the receive operator parallelism and the distribution being used. At worse, there is one mailbox per worker pair, so if the upstream send operator has a parallelism of S and the receive operator has a parallelism of R, there will be S * R mailboxes.

The hash join operator is used to join two relations using a hash join algorithm. It is a binary operator that takes two inputs, the left and right relations, and produces a single output relation.

This is the only join operator in the multi-stage query engine and it is always created as a result of a query that contains a join clause, but can be created by other SQL queries like ones using semi-join.

There are different types of joins that can be performed using the hash join operator. Apache Pinot supports:

• Inner join, where only the rows that have a match in both relations are returned.

• Left join, where all the rows from the left relation are returned. The ones that have a match with the right relation are returned with the columns from the right relation, and the ones that do not have a match are returned with null values for the columns from the right relation.

• Right join, like the left join but returning all the rows from the right relation, with the columns from the left relation filled with null values for the rows that do not have a match.

• Full outer join, where all the rows from both relations are returned. If a row from any relation does not have a match in the other relation, the columns from the other relation are filled with null values.

• Semi-join, where only the rows from the left relation that have a match in the right relation are returned. This is useful to filter the rows from the left relation based on the existence of a match in the right relation.

• Anti-join, where only the rows from the left relation that do not have a match in the right relation are returned.

Implementation details

The hash join operator is one of the new operators introduced in the multi-stage query engine. The current implementation assumes that the right input relation is the smaller one, so it consumes this input first building a hash table that is then probed with the left input relation.

Although the whole multi-stage query engine is designed to be able to process the data in memory, the multi-stage query engine uses the ability to execute each stage in different workers (explained in ) to be able to process the data that may not fit in the memory of a single node. Specifically, each worker processes a subset of the data. Inputs are by default partitioned by the join keys and each worker process one partition of the data.

This means that data usually needs to be shuffled between workers, which is done by the engine using a mailbox system. The engine tries to minimize the amount of data that needs to be shuffled by partitioning the data, but some techniques can be used to reduce the amount of data that needs to be shuffled, like using co-located joins.

Blocking nature

The hash join operator is a blocking operator. It needs to consume all the input data (from both inputs) before emitting the result.

Maximum number of groups

Even using partitioning, the amount of data that needs to be stored in memory can be high, so the engine tries to protect itself from running out of memory by limiting the number of groups that can be created by the join keys.

The hint can be used to control the behavior of the engine when the number of groups exceeds the limit. This limit can be defined using the hint. By default, this limit is slightly above 1 million groups and the default join overflow mode is THROW, which means that the query will fail if the number of groups exceeds the limit.

Hints

join_overflow_mode

Type: String

Default: THROW

Defines the behavior of the engine when the number of groups exceeds the limit defined by the max_rows_in_join hint. The possible values are:

• THROW: The query will fail if the number of groups exceeds the limit.

• BREAK: The engine will stop processing the join and return the results that have been computed so far. In this case the stat maxRowsInJoinReached will be true.

max_rows_in_join

Type: Integer

Default: 1.048.576

The maximum number of groups that can be created by the join keys. What happens when this limit is reached is defined by the join_overflow_mode hint.

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1.

emittedRows

Type: Long

The number of groups emitted by the operator. Joins can emit more rows than the input relations, so this value can be higher than the number of rows in the input. Remember that the number of groups is limited by the max_rows_in_join hint and a large number of groups can lead to high memory usage and long GC pauses, which can affect the performance of the whole system.

maxRowsInJoinReached

Type: Boolean

This stat is set to true when the number of groups exceeds the limit defined by the max_rows_in_join hint.

Notice that by default the engine will throw an exception when this happens in which case no stat will be emitted. Therefore this stat is only emitted when the join_overflow_mode hint is set to BREAK.

timeBuildingHashTableMs

Type: Long

The time spent building the hash table used to probe the join keys, in milliseconds.

A large number here can indicate that the right relation is too large or the right relation is taking too long to be processed.

Explain attributes

The hash join operator is represented in the explain plan as a LogicalJoin explain node.

condition

Type: Expression

The condition that is being applied to the rows to join the relations. The expression may use indexed columns ($0, $1, etc), functions and literals. The indexed columns are always 0-based.

For example, the following explain plan:

Is saying that the join condition is that the column with index 0 in the left relation is equal to the column with index 1 in the right relation. Given the rest of the explain plan, we can see that the column with index 0 userUUID column in the userAttributes table and the column with index 1 is the userUUID column in the userGroups table.

joinType

Type: String

The type of join that is being performed. The possible values are: inner, left, right, full, semi and anti, as explained in .

Tips and tricks

The order of input relations matter

Apache Pinot does not use table stats to determine the best order to consume the input relations. Instead, it assumes that the right input relation is the smaller one. That relation will always be fully consumed to build a hash table and sometimes it will be broadcasted to all workers. This means that it is important to specify the smaller relation as the right input.

Remember that left and right are relative to the order of the tables in the SQL query. It is less expensive to do a join between a large table and a small table than the other way around.

For example, this query:

is more efficient than:

The window operator is used to define a window over which to perform calculations.

This page describes the window operator defined in the relational algebra used by multi-stage queries. This operator is generated by the multi-stage query engine when you window functions in a query. You can read more about window functions in the windows functions reference documentation.

Unlike the aggregate operator, which will output one row per group, the window operator will output as many rows as input rows.

Implementation details

Window operators take a single input relation and apply window functions to it. For each input row, a window of rows is calculated and one or many aggregations are applied to it.

In general window operator are expensive in terms of CPU and memory usage, but they open the door to a wide range of analytical queries.

Blocking nature

The window operator is a blocking operator. It needs to consume all the input data before emitting the result.

Hints

Window hints are configured with the windowOptions hint, which accepts as argument a map of options and values.

For example:

max_rows_in_window

Type: Integer

Default: 1048576

Max rows allowed to cache the rows in window for further processing.

window_overflow_mode

Type: THROW or BREAK

Default: 'THROW'

Mode when window overflow happens, supported values:

• THROW: Break window cache build process, and throw exception, no further WINDOW operation performed.

• BREAK: Break window cache build process, continue to perform WINDOW operation, results might be partial.

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1. This number is affected by the number of received rows and the complexity of the window function.

emittedRows

Type: Long

The number of groups emitted by the operator. A large number of emitted rows can indicate that the query is not well optimized.

Unlike the , which will output one row per group, the window operator will output as many rows as input rows.

maxRowsInWindowReached

Type: Boolean

This attribute is set to true if the maximum number of rows in the window has been reached.

Explain attributes

The window operator is represented in the explain plan as a LogicalWindow explain node.

window#

Type: Expression

The window expressions used by the operator. There may be more than one of these attributes depending on the number of window functions used in the query, although sometimes multiple window function clauses in SQL can be combined into a single window operator.

The expression may use indexed columns ($0, $1, etc) that represent the columns of the virtual row generated by the upstream.

Tips and tricks

None

The union operator combines the results of two or more queries into a single result set. The result set contains all the rows from the queries. Contrary to other set operations ( and ), the union operator does not remove duplicates from the result set. Therefore its semantic is similar to the SQL UNION or UNION ALL operator.

There is no guarantee on the order of the rows in the result set.

The transform operator is used to apply a transformation to the input data. They may filter out columns or add new ones by applying functions to the existing columns. This operator is generated by the multi-stage query engine when you use a SELECT clause in a query, but can also be used to implement other transformations.

Implementation details

Transform operators apply some transformation functions to the input data received from upstream. The cost of the transformation usually depends on the complexity of the functions applied, but comparing to other operators, it is usually not very high.

The minus operator is used to subtract the result of one query from another query. This operator is used to find the difference between two sets of rows, usually by using the SQL EXCEPT operator.

Implementation details

The minus operator is a semi-blocking operator that first consumes the right input relation in a blocking fashion and then consumes the left input relation in a streaming fashion.

The current implementation consumes the whole right input relation first and stores the rows in a set. Then it consumes the left input relation one block at a time. Each time a block of rows is read from the left input relation, the operator checks if the rows are in the set of rows from the right input relation. All unique rows that are not in the set are added to a new partial result block. Once the whole left input block is analyzed, the operator emits the partial result block.

This process is repeated until all rows from the left input relation are processed.

Blocking nature

The minus operator is a semi-blocking operator that first consumes the right input relation in a blocking fashion and then consumes the left input relation in a streaming fashion.

In pseudo-code, the algorithm looks like this:

Hints

None

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1.

emittedRows

Type: Long

The number of groups emitted by the operator.

Explain attributes

The minus operator is represented in the explain plan as a LogicalMinus explain node.

all

Type: Boolean

This attribute is used to indicate if the operator should return all the rows or only the distinct rows.

Tips and tricks

Memory pressure

The minus operator ends up having to store all unique rows from both input relations in memory. This can lead to memory pressure if the input relations are large and have a high number of unique rows.

The order of input relations matter

Although the minus operator ends up adding all unique rows from both input relations to a set, the order of input relations matters. While the right input relation is consumed in a blocking fashion, the left input relation is consumed in a streaming fashion. Therefore the latency of the whole query could be improved if the left input relation is producing values in streaming fashion.

The intersect operator is a relational operator that combines two relations and returns the common rows between them. The operator is used to find the intersection of two or more relations, usually by using the SQL INTERSECT operator.

Implementation details

The current implementation consumes the whole right input relation first and stores the rows in a set. Then it consumes the left input relation one block at a time. Each time a block of rows is read from the left input relation, the operator checks if the rows are in the set of rows from the right input relation. All unique rows that are in the set are added to a new partial result block. Once the whole left input block is analyzed, the operator emits the partial result block.

This process is repeated until all rows from the left input relation are processed.

In pseudo-code, the algorithm looks like this:

Blocking nature

The intersect operator is a semi-blocking operator that first consumes the right input relation in a blocking fashion and then consumes the left input relation in a streaming fashion.

Hints

None

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1.

emittedRows

Type: Long

The number of groups emitted by the operator.

Explain attributes

The intersect operator is represented in the explain plan as a LogicalIntersect explain node.

all

Type: Boolean

This attribute is used to indicate if the operator should return all the rows or only the distinct rows.

Tips and tricks

The order of input relations matter

The intersect operator has a memory footprint that is proportional to the number of unique rows in the right input relation. It also consumes the right input relation in a blocking fashion while the left input relation is consumed in a streaming fashion.

This means that:

• In case any of the input relations is significantly larger than the other, it is recommended to use the smaller relation as the right input relation.

• In case one of the input is blocking and the other is not, it is recommended to use the blocking relation as the right input relation.

These two hints can be contradictory, so it is up to the user to decide which one to follow based on the specific query pattern. Remember that you can use the stage stats to check the number of rows emitted by each of the inputs and adjust the order of the inputs accordingly.

Deep dive into stages

As explained in the Multi-stage query engine reference documentation, the multi-stage query engine breaks down a query into multiple stages. Each stage corresponds to a subset of the query plan and is executed independently. Stages are connected in a tree-like structure where the output of one stage is the input to another stage. The stage that is at the root of the tree sends the final results to the client. The stages that are at the leaves of the tree read from the tables. The intermediate stages process the data and send it to the next stage.

When the broker receives a query, it generates a query plan. This is a tree-like structure where each node is an operator. The plan is then optimized, moving and changing nodes to generate a plan that is semantically equivalent (it returns the same rows) but more efficient. During this phase the broker colors the nodes of the plan, assigning them to a stage. The broker also assigns a parallelism to each stage and defines which servers are going to execute each stage. For example, if a stage has a parallelism of 10, then at most 10 servers will execute that stage in parallel. One single server can execute multiple stages in parallel and it can even execute multiple instances of the same stage in parallel.

Stages are identified by their stage ID, which is a unique identifier for each stage. In the current implementation the stage ID is a number and the root stage has a stage ID of 0, although this may change in the future.

The current implementation has some properties that are worth mentioning:

• The leaf stages execute a slightly modified version of the single-stage query engine. Therefore these stages cannot execute joins or aggregations, which are always executed in the intermediate stages.

• Intermediate stages execute operations using a new query execution engine that has been created for the multi-stage query engine. This is why some of the functions that are supported in the single-stage query engine are not supported in the multi-stage query engine and vice versa.

• An intermediate stage can only have one join, one window function or one set operation. If a query has more than one of these operations, the broker will create multiple stages, each with one of these operations.

Extracting Stages from Query Plans

As explained in , you can use the EXPLAIN PLAN syntax to obtain the logical plan of a query. This logical plan can be used to extract the stages of the query.

For example, if the query is:

A possible output of the EXPLAIN PLAN command is:

As it happens with all queries, the logical plan forms a tree-like structure. In this default explain format, the tree-like structure is represented with indentation. The root of the tree is the first line, which is the last operator to be executed and marks the root stage. The boundary between stages are the PinotLogicalExchange operators. In the example above, there are four stages:

• The root stage starts with the LogicalSort operator in the root of operators and ends with the PinotLogicalSortExchange operator. This is the last stage to be executed and the only one that is executed in the broker, which will directly send the result to the client once it is computed.

• The next stage starts with this PinotLogicalSortExchange operator and includes the LogicalSort operator, the LogicalProject operator, the LogicalJoin

Now that we have identified the stages, we can understand what each stage is doing by .

Multi-stage plans are a bit more complex than single-stage plans. This page explains how to interpret multi-stage explain plans.

As explained in , you can use the EXPLAIN PLAN syntax to obtain the logical plan of a query. There are different formats for the output of the EXPLAIN PLAN command, but all of them represent the logical plan of the query.

The query

Can produce the following output:

We can see that each node in the tree represents an operation that is executed in the query and each operator has some attributes. For example the LogicalJoin

The leaf operator is the operator that actually reads the data from the segments. Instead of being just a simple table scan, the leaf operator is a meta-operator that wraps the single-stage query engine and executes all the operators in the leaf stage of the query plan.

Implementation details

The leaf operator is not a relational operator itself but a meta-operator that is able to execute single-stage queries. When servers execute a leaf stage, they compile all operations in the stage but the send operator into the equivalent single-stage query and execute that using a slightly modified version of the single-stage engine.

As a result, leaf stage operators can use all the optimizations and indices that the single-stage engine can use but it also means that there may be slight differences when an operator is executed in a leaf stage compared to when it is executed in an intermediate stage. For example, operations pushed down to the leaf stage may use indexes (see ) or the semantics can be slightly different.

You can read for more information on the differences between the leaf and intermediate stages, but the main ones are:

• Null handling is different.

• Some functions are only supported in multi-stage and some others only in single-stage.

• Type coercion is different. While the single-stage engine always operates with generic types (ie uses doubles when mathematical operations are used), the multi-stage engine tries to keep the types (ie adding two integers will result in an integer).

Blocking nature

One of the slight differences between the leaf and the normal single-stage engine is that the leaf engine tries to be not blocking.

Hints

None

Stats

executionTimeMs

Type: Long

The summation of time spent by all threads executing the operator. This means that the wall time spent in the operation may be smaller that this value if the parallelism is larger than 1.

emittedRows

Type: Long

The number of groups emitted by the operator.

table

Type: String

The name of the table that is scanned. This is the name without the type suffix (so without _REALTIME or _OFFLINE). This is very useful to understand which table is being scanned by this leaf stage in case of complex queries.

numDocsScanned

Type: Long

Similar to the same stat in single-stage queries, this stat indicates the number of rows selected after the filter phase.

If it is very high, that means the selectivity for the query is low and lots of rows need to be processed after the filtering. Consider refining the filter to increase the selectivity of the query.

numEntriesScannedInFilter

Type: Long

Similar to the same stat in single-stage queries, this stat indicates the number of entries (aka scalar values) scanned in the filtering phase of query execution.

Can be larger than the total scanned doc count because of multiple filtering predicates or multi-value entries. Can also be smaller than the total scanned doc count if indexing is used for filtering.

This along with numEntriesScannedPostFilter indicates where most of the time is spent during table scan processing. If this value is high, enabling indexing for affected columns is a way to bring it down. Another option is to partition the data based on the dimension most heavily used in your filter queries.

numEntriesScannedPostFilter

Type: Long

Similar to the same stat in single-stage queries, this stat indicates the number of entries (aka scalar values) scanned after the filtering phase of query execution, ie. aggregation and/or group-by phases. This is equivalent to numDocScanned * number of projected columns.

This along with numEntriesScannedInFilter indicates where most of the time is spent during table scan processing. A high number for this means the selectivity is low (that is, Pinot needs to scan a lot of records to answer the query). If this is high, consider using star-tree index, given a regular index won't improve performance.

numSegmentsQueried

Type: Integer

Similar to the same stat in single-stage queries, this stat indicates the total number of segment queried for a query. May be less than the total number of segments if the broker applies optimizations.

The broker decides how many segments to query on each server, based on broker pruning logic. The server decides how many of these segments to actually look at, based on server pruning logic. After processing segments for a query, fewer may have the matching records.

In general, numSegmentsQueried >= numSegmentsProcessed >= numSegmentsMatched.

numSegmentsProcessed

Type: Integer

Similar to the same stat in single-stage queries, this stat indicates the number of segments processed with at least one document matched in the query response.

The more segments are processed, the more IO has to be done. This is why selective queries where numSegmentsProcessed is close to numSegmentsQueried can be optimized by changing the data distribution.

numSegmentsMatched

Type: Integer

Similar to the same stat in single-stage queries, this stat indicates the number of segment operators used to process segments. Indicates the effectiveness of the pruning logic.

totalDocs

Type: Long

Similar to the same stat in single-stage queries, this stat indicates the number of rows in the table.

numGroupsLimitReached

Type: Boolean

Similar to the same stat in single-stage queries and the same in operators, this stat indicates if the max group limit has been reached in a group by aggregation operator executed in the leaf stage.

If this boolean is set to true, the query result may not be accurate. The default value for numGroupsLimit is 100k, and should be sufficient for most use cases.

numResizes

Type: Integer

Number of result resizes for queries

resizeTimeMs

Type: Long

Time spent in resizing results for the output. Either because of LIMIT or maximum allowed group by keys or any other criteria.

threadCpuTimeNs

Type: Long

Aggregated thread cpu time in nanoseconds for query processing from servers. This metric is only available if Pinot is configured with pinot.server.instance.enableThreadCpuTimeMeasurement.

systemActivitiesCpuTimeNs

Type: Long

Aggregated system activities cpu time in nanoseconds for query processing (e.g. GC, OS paging etc.) This metric is only available if Pinot is configured with pinot.server.instance.enableThreadCpuTimeMeasurement.

numSegmentsPrunedByServer

Type: Integer

The number of segments pruned by the server, for any reason.

numSegmentsPrunedInvalid

Type: Integer

The number of segments pruned because they are invalid. Segments are invalid when the schema has changed and the segment has not been refreshed.

For example, if a column is added to the schema, the segment will be invalid for queries that use that column until it is refreshed.

numSegmentsPrunedByLimit

Type: Integer

The number of segments pruned because they are not needed for the query due to the limit clause.

Pinot keeps a count of the number of rows returned by each segment. Once it's guaranteed that no more segments need to be read to satisfy the limit clause without breaking semantics, the remaining segments are pruned.

For example, a query like SELECT col1 FROM table2 LIMIT 10 can be pruned for this reason while a query like SELECT col1 FROM table2 ORDER BY col1 DESC LIMIT 10 cannot because Pinot needs to read all segments to guarantee the larger values of col1 are returned.

numSegmentsPrunedByValue

Type: Integer

The number of segments pruned because they are not needed for the query due to a value clause, usually a where.

Pinot keeps the maximum and minimum values of each segment for each column. If the value clause is such that the segment cannot contain any rows that satisfy the clause, the segment is pruned.

numConsumingSegmentsProcessed

Type: Integer

Like numSegmentsProcessed but only for consuming segments.

numConsumingSegmentsMatched

Type: Integer

Like numSegmentsMatched but only for consuming segments.

operatorExecutionTimeMs

Type: Long

The time spent by the operator executing.

operatorExecStartTimeMs

Type: Long

The instant in time when the operator started executing.

Explain attributes

Given that the leaf operator is a meta-operator, it is not actually shown in the explain plan. But the leaf stage is the only operator that can execute table scans, so here we list the attributes that can be found in the explain plan for a table scan

table

Type: String array

Example: table=[[default, userGroups]]

The qualified name of the table that is scanned, which means it also contains the name of the database being used.

Tips and tricks

Try to push as much as possible to the leaf stage

Leaf stage operators can use all the optimizations and indices that the single-stage engine can use. This means that it is usually better to push down as much as possible to the leaf stage.

The engine is smart enough to push down filters and aggregations without breaking semantics, but sometimes there are subtle SQL semantics and what the domain expert writing the query wants to do.

Sometimes things the engine is too paranoid about null handling or the query includes an unnecessary limit clause that prevents the engine from pushing down the filter.

It is recommended to analyze your explain plan to be sure that the engine is able to push down as much logic as you expect.

The sort or limit operator is used to sort the input data, limit the number of rows emitted by the operator or both. This operator is generated by the multi-stage query engine when you use an order by, limit or offset operation in a query.

Implementation details

Multi-stage stats are more complex but also more expressive than single-stage stats. While in single-stage stats Apache Pinot returns a single set of statistics for the query, in multi-stage stats Apache Pinot returns a set of statistics for each operator of the query execution.

These stats can be seen when using Pinot controller UI by running the query and clicking on the Show JSON format button. Then the whole JSON response will be shown and the multi-stage stats will be in a field called stageStats. Different drivers may provide different ways to see the stats.

For example the following query:

SELECT playerName, teamName
FROM baseballStats_OFFLINE as playerStats
JOIN dimBaseballTeams_OFFLINE AS teams
    ON playerStats.teamID = teams.teamID
LIMIT 10

Returns the following stageStats:

Each node in the tree represents an operation that is executed and the tree structure form is similar (but not equal) to the logical plan of the query that can be obtained with the EXPLAIN PLAN command.

As you can see, each operator has a type and the stats carried on the node depend on that type. You can learn more about each operator types and their stats in the section.