8.5. Spill to Disk#
In the case of memory intensive operations, Presto allows offloading intermediate operation results to disk. The goal of this mechanism is to enable execution of queries that require amounts of memory exceeding per query or per node limits.
The mechanism is similar to OS level page swapping. However, it is implemented on the application level to address specific needs of Presto.
Properties related to spilling are described in Spilling Properties.
Memory Management and Spill#
By default, Presto kills queries, if the memory requested by the query execution
exceeds session properties
query_max_memory_per_node. This mechanism ensures fairness in allocation
of memory to queries, and prevents deadlock caused by memory allocation.
It is efficient when there is a lot of small queries in the cluster, but
leads to killing large queries that don’t stay within the limits.
To overcome this inefficiency, the concept of revocable memory was introduced. A query can request memory that does not count toward the limits, but this memory can be revoked by the memory manager at any time. When memory is revoked, the query runner spills intermediate data from memory to disk and continues to process it later.
In practice, when the cluster is idle, and all memory is available, a memory intensive query may use all of the memory in the cluster. On the other hand, when the cluster does not have much free memory, the same query may be forced to use disk as storage for intermediate data. A query, that is forced to spill to disk, may have a longer execution time by orders of magnitude than a query that runs completely in memory.
Please note that enabling spill-to-disk does not guarantee execution of all
memory intensive queries. It is still possible that the query runner fails
to divide intermediate data into chunks small enough so that every chunk fits into
memory, leading to
Out of memory errors while loading the data from disk.
Spill Disk Space#
Spilling intermediate results to disk, and retrieving them back, is expensive
in terms of IO operations. Thus, queries that use spill likely become
throttled by disk. To increase query performance, it is recommended to
provide multiple paths on separate local devices for spill (property
spiller-spill-path in Spilling Properties).
The system drive should not be used for spilling, especially not to the drive where the JVM is running and writing logs. Doing so may lead to cluster instability. Additionally, it is recommended to monitor the disk saturation of the configured spill paths.
Presto treats spill paths as independent disks (see JBOD), so there is no need to use RAID for spill.
When spill compression is enabled (
spill-compression-enabled property in
Spilling Properties), spilled pages are compressed, before being
written to disk. Enabling this feature can reduce disk IO at the cost
of extra CPU load to compress and decompress spilled pages.
When spill encryption is enabled (
spill-encryption-enabled property in
Spilling Properties), spill contents are encrypted with a randomly generated
(per spill file) secret key. Enabling this increases CPU load and reduces throughput
of spilling to disk, but can protect spilled data from being recovered from spill files.
Consider reducing the value of
memory-revoking-threshold when spill
encryption is enabled, to account for the increase in latency of spilling.
Not all operations support spilling to disk, and each handles spilling differently. Currently, the mechanism is implemented for the following operations.
During the join operation, one of the tables being joined is stored in memory. This table is called the build table. The rows from the other table stream through and are passed onto the next operation, if they match rows in the build table. The most memory-intensive part of the join is this build table.
When the task concurrency is greater than one, the build table is partitioned.
The number of partitions is equal to the value of the
configuration parameter (see Task Properties).
When the build table is partitioned, the spill-to-disk mechanism can decrease the peak memory usage needed by the join operation. When a query approaches the memory limit, a subset of the partitions of the build table gets spilled to disk, along with rows from the other table that fall into those same partitions. The number of partitions, that get spilled, influences the amount of disk space needed.
Afterward, the spilled partitions are read back one-by-one to finish the join operation.
With this mechanism, the peak memory used by the join operator can be decreased
to the size of the largest build table partition. Assuming no data skew, this
1 / task.concurrency times the size of the whole build table.
Aggregation functions perform an operation on a group of values and return one value. If the number of groups you’re aggregating over is large, a significant amount of memory may be needed. When spill-to-disk is enabled, if there is not enough memory, intermediate cumulated aggregation results are written to disk. They are loaded back and merged with a lower memory footprint.
If your trying to sort a larger amount of data, a significant amount of memory
may be needed. When spill to disk for
order by is enabled, if there is not enough
memory, intermediate sorted results are written to disk. They are loaded back and
merged with a lower memory footprint.
Window functions perform an operator over a window of rows, and return one value for each row. If this window of rows is large, a significant amount of memory may be needed. When spill to disk for window functions is enabled, if there is not enough memory, intermediate sorted results are written to disk. They are loaded back and merged when memory is available. There is a current limitation that spill does not work in all cases, such as when a single window is very large.