Parallelizing queries

Deephaven supports using multiple processors to speed up query evaluation. The extent to which Deephaven employs multiple processors depends on both the phase of operation and the query itself.

Query initialization and updates

When considering Deephaven query performance, there are two distinct phases to consider: initialization and updates.

Query initialization

Every table operation method — .where, .update, .naturalJoin, etc. — undergoes an initialization phase when the method is called. Initialization produces a result table based on the data in the source table. For example, with a 100,000-row table called myTable, running myTable.update("X = random()") will run the random() method 100,000 times (once per row).

If an operation's source table is refreshing, then initialization will create a new node in the update graph as well.

Query updates

After initialization, the Periodic Update Graph (UG) monitors source tables for changes and process updates to any table. For example, if 25,000 rows are added to myTable, the UG will run the random() method 25,000 more times, calculating the value of column X for each of the new rows.

Parallelizing queries

Parallelizing query initialization

Deephaven is a column-oriented query engine — it focuses on handling data one column at a time, instead of one row at a time like many traditional databases. Since Deephaven column sources support random access to data, different segments of a column can be processed in parallel. When possible, the Deephaven engine will do this automatically, based on the number of threads in the Operation Initialization Thread Pool.

Parallelizing query updates

As with query initialization, some operations can process different sections of a column in parallel. However, update processing can also be parallelized across independent nodes of the DAG. Parallel processing of updates depends on the size of the Periodic Update Graph Thread Pool.

Consider the following hypothetical example:

// Retrieve a live table:
my_table = get_my_kafka_feed_table()

// Run several independent query operations on 'my_table':
my_table_updated = my_table.update("MyCalculation = computeValue(Col1, Col2, ColRed, ColBlue)")
my_table_filtered1 = my_table.where("ColX < 10000")
my_table_filtered2 = my_table.where("ColY > ColZ")

// Create a result table that depends on the three prior tables:
merged_tables = merge(
    my_table_updated,
    my_table_filtered1,
    my_table_filtered2
)

The three intermediate tables my_table_updated, my_table_filtered1 and my_table_filtered2 all depend on only one other table — the original my_table. Since they are independent of each other, when my_table is updated with new or modified rows it is possible for the query engine to process the new rows into my_table_updated, my_table_filtered1 and my_table_filtered2 at the same time. However, since merged_tables depends on those three tables, the query engine cannot update the result of the merge operation until after the update() and where()s for those three tables have been processed.

Controlling Concurrency for select, update and where

The select, update, and where operations can parallelize within a single where clause or column expression. This can greatly improve throughput by using multiple threads to read existing columns or compute functions. Deephaven can only parallelize an expression if it is stateless, meaning it does not depend on any mutable external inputs or the order in which rows are evaluated. Many operations, such as String manipulation or arithmetic on one or more input columns are stateless. By default, the Deephaven engine assumes that expressions are not stateless. For select and update, you can change the configuration property QueryTable.statelessSelectByDefault to true to make columns stateless by default. For filters, change the property QueryTable.statelessFiltersByDefault.

The ConcurrencyControl interface allows you to control the behavior of Filter (where clause) and Selectable (column formula) objects.

ConccurencyControl cannot be applied to Selectables passed to view or updateView. The view and updateView operations compute results on demand, and therefore cannot enforce ordering constraints.

To explicitly mark a Selectable or Filter as stateful, use the withSerial method.

• A serial Filter cannot be reordered with respect to other Filters. Every input row to a stateful Filter is evaluated in order.
• When a Selectable is serial, then every row for that column is evaluated in order.
• For Selectables, additional ordering constraints are controlled by the value of the QueryTable.SERIAL_SELECT_IMPLICIT_BARRIERS. This is set by the property QueryTable.serialSelectImplicitBarriers. The default value is the inverse of QueryTable.statelessSelectByDefault. When Selectables are stateless by default, no implicit barriers are added (i.e., QueryTable.SERIAL_SELECT_IMPLICIT_BARRIERS is false). When Selectables are stateful by default, then implicit barriers are added (i.e. QueryTable.SERIAL_SELECT_IMPLICIT_BARRIERS is true).
• If QueryTable.SERIAL_SELECT_IMPLICIT_BARRIERS is false, no additional ordering between expressions is imposed. As with every select or update call, if column B references column A, then the necessary inputs to column B from column A are evaluated before column B is evaluated. To impose further ordering constraints, use barriers.
• If QueryTable.SERIAL_SELECT_IMPLICIT_BARRIERS is true, then a serial selectable is an absolute barrier with respect to all other serial selectables. This prohibits serial selectables from being evaluated concurrently, permitting them to access global state. Selectables that are not serial may be reordered with respect to a serial selectable.

Filters and Selectables may declare a barrier. A barrier is an opaque object (compared using reference equality) that is used to mark a particular Filter or Selectable. Subsequent Filters or Selectables may respect a previously declared barrier. If a Filter respects a barrier, that Filter cannot begin evaluation until the Filter which declares the barrier has been completely evaluated. Similarly, if a Selectable respects a barrier, then it cannot begin evaluation until the Selectable which declared the barrier has been completely evaluated.

In this code block, two columns reference the AtomicInteger a:

import java.util.concurrent.atomic.AtomicInteger

a = new AtomicInteger(0)
t = emptyTable(1_000_000).update("A=a.getAndIncrement()", "B=a.getAndIncrement()")

Deephaven's default behavior is to treat both A and B statefully, therefore the table is equivalent to:

t = emptyTable(1_000_000).update("A=i", "B=1_000_000 + i")

However, when the columns are stateless, then the rows from either column can be evaluated in any order. To indicate that A must be evaluated before B, we can use a barrier:

import java.util.concurrent.atomic.AtomicInteger

a = new AtomicInteger(0)
t = emptyTable(1_000_000).update(List.of(
        Selectable.of(ColumnName.of("A"), RawString.of("a.getAndIncrement()")).withDeclaredBarriers(a),
        Selectable.of(ColumnName.of("B"), RawString.of("a.getAndIncrement()")).withRespectedBarriers(a)))

Similarly, we can prevent values of A from appearing out of order using withSerial:

import java.util.concurrent.atomic.AtomicInteger

a = new AtomicInteger(0)
t=emptyTable(1_000_000).update(List.of(
        Selectable.of(ColumnName.of("A"), RawString.of("a.getAndIncrement()")).withSerial(),
        Selectable.of(ColumnName.of("B"), RawString.of("a.getAndIncrement()"))))

Managing thread pool sizes

The maximum parallelism of query initialization and update processing is determined by the Operation Initialization Thread Pool and the Periodic Update Graph Thread Pool. The size of these values is controlled using the properties described in the table below:

Setting either of these properties to -1 instructs Deephaven to use all available processors. The number of available processors is retrieved from the Java Virtual Machine at Deephaven startup, using Runtime.availableProcessors().

Related documentation

• Deephaven’s Directed-Acyclic-Graph (DAG)
• Multithreading: Synchronization, locks, and snapshots