Skip to content

Apache DataFusion SQL Query Engine

28.9 KB
Raw Blame

SQL Benchmarks

This directory contains a collection of benchmarks each driven by a simple '.benchmark' text file and sql queries that exercise the DataFusion execution engine against a variety of benchmark suites. The sql benchmark framework is intentionally simple so that benchmarks and queries can be added or modified without touching the core engine or requiring recompilation.

The sql benchmarks are organized in sub‑directories that correspond to the benchmark suites that are commonly used in the community:

Benchmark Suite Description
asof_join ASOF join benchmarks across size, ordering, grouping, match direction, and payload width
clickbench ClickBench benchmark
clickbench extended 12 additional, more complex queries against the Clickbench dataset
clickbench_sorted ClickBench benchmark using a pre-sorted hits file.
h2o The h2o benchmark
hj Hash join benchmark
imdb IMDb benchmark
nlj Nested‑loop join benchmark
null_aware_join Null-aware (NOT IN) hash join micro-benchmarks. Q01-Q03 are uncorrelated NOT IN across NULL fractions and are linear in the table size (NAJ_LARGE_ROWS, default 1000000). Q04-Q08 are correlated, so the correlation predicate stays behind as a join filter that the join applies per candidate (build row × probe row) pair while deciding which outer rows are UNKNOWN; without an equality correlation there are no scope keys to narrow those pairs, so their cost grows with the square of NAJ_ROWS (default 10000). Q08 adds an equality correlation, which turns those pairs into a hash lookup. All tables are built inline from range(), so there is no data step.
push_down_topk ORDER BY ... LIMIT over outer joins (TPC-H data); exercises pushing a TopK through a join
pwmj Piecewise merge join: classic INNER/LEFT/RIGHT/FULL joins over lhs.key < rhs.key. Subgroups (--subgroup) are match regimes: no_match, null_heavy, half_match, selective, all_match. Generated with range(), no data needed.
smj Sort‑merge join benchmark
sort_pushdown SQL counterparts of the five sort pushdown variants, using existing Parquet data and query files.
sort tpch Sorting benchmarks against the TPC-H lineitem table
taxi NYC taxi dataset benchmark
tpcds TPC‑DS queries
tpch TPC‑H queries
wide_schema Small-projection queries on a wide (1024-col, 256-file) synthetic dataset; runs wide + narrow subgroups for comparison
predicate_eval Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an adaptive predicate-ordering system behaves across them (#11262). Subgroups (--subgroup): costsel, cost, selectivity, cardinality, width, scale, neutral, correlation, drift, nulls. The suite sets no engine config of its own, so by default it measures DataFusion's built-in left-deep AND short-circuit; point it at a system under test by exporting that system's own DataFusion setting (the harness builds its SessionConfig with SessionConfig::from_env). Every query is a count(*), and the counts are checked in under predicate_eval/results/, so --result-mode validate also checks that a reordering under test still returns the same rows; the checked-in counts were persisted at the suite defaults (PRED_ROWS=1000000, PRED_FILL=30), so validation assumes those (the scale and width subgroups pin their own values per query and validate at any setting).
parquet_row_filter_skip Micro-benchmark for the per-row-group fully-matched RowFilter skip on Parquet scans (#23696). Subgroups (--subgroup): skip (clustered key, most row groups fully matched by statistics so the per-row filter is skipped), control (scrambled key, no row group is ever fully matched). Size the data with PRED_ROWS and the row-group size with RG_SIZE.
spill_views Sort and GROUP BY queries that spill StringView and BinaryView columns read from Parquet (#23564). Each query sets its own memory limit and target_partitions = 4, so it spills in any environment. Subgroups (--subgroup): repeated (sort with a low-cardinality payload that shares one dictionary buffer, where spill compaction can remove data), distinct (sort and GROUP BY on all-distinct strings, where compaction cannot remove data). Set the memory limit of each subgroup with SPILL_VIEWS_LIMIT_REPEATED (default 40M) and SPILL_VIEWS_LIMIT_DISTINCT (default 96M), for example to let both sides of a comparison finish when one of them runs out of memory at the default. The queries must still spill, so check a raised limit with EXPLAIN ANALYZE and keep it well below the working set: q05 stops spilling at about 192M. Give a whole number of K, M or G units, because the run asserts the limit that the memory pool reports back.
projection_subquery IN, NOT IN and EXISTS subqueries that sit in the SELECT list instead of a filter (#25341). Every query projects the boolean subquery result and aggregates it, so the measured cost is the decorrelation plan and not the size of the output. q07 correlates on < instead of =, so the correlation stays a residual filter of the mark join. Size the two inline tables with PSQ_ROWS (default 30000). The counts are checked in under projection_subquery/results/, so --result-mode validate also proves that a change to the decorrelation still returns the same rows; the checked-in counts were persisted at the default PSQ_ROWS, so validation assumes that value.

Running Benchmarks

Use benchmark_runner to run SQL benchmarks. It reads each suite's .suite file and exposes the suite's configuration as command-line options. Use the bench.sh shell script one level above this directory to download or generate required data files.

cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- tpch

SQL benchmark runner

The benchmark_runner binary discovers suites from this directory and exposes suite-specific options alongside the common benchmark options. The suite name must come before all options.

cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- --list
cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- tpch --help
cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- tpch --query 15 --format csv
cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- clickbench --partitioning partitioned --dry-run
cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- tpch --query 1 --result-mode persist
cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- tpch --query 1 --result-mode validate

Use --path PATH or -p PATH to override DATA_DIR for a suite that declares that path replacement. Suites without a DATA_DIR replacement reject the option. Suite-specific values follow this precedence: command-line option, environment variable, then the default in the suite metadata.

--dry-run prints the resolved suite, filters, run mode, common options, suite-specific values, and path replacements as JSON. It reports source metadata for suite options and path replacements. It validates the command but does not load benchmark definitions, create a session, read datasets, execute SQL, or write benchmark results.

Use --result-mode persist to save query results or --result-mode validate to compare them with saved results. The default, --result-mode none, does neither. For compatibility with direct Criterion runs, the runner also reads BENCH_PERSIST_RESULTS and BENCH_VALIDATE. Persistence takes precedence when both variables are true. An explicit --result-mode overrides both variables.

Suite metadata

Each discoverable suite has one TOML metadata file named <suite>/<suite>.suite. The runner accepts these top-level fields:

Field Required Description
description Yes Non-empty text shown by --list and suite help.
query_pattern No Relative benchmark filename pattern. It must contain exactly one {QUERY_ID} or {QUERY_ID_PADDED} placeholder and defaults to q{QUERY_ID_PADDED}.benchmark.
path_replacements No Map of replacement names to paths. Relative paths resolve from the suite directory. DATA_DIR enables --path/-p.
options No Array of suite-specific option tables described below.
examples No Array of command and description pairs appended to suite help. Both values must contain text.

Each [[options]] table has these fields:

Field Required Description
name Yes Long option name without --; use lowercase ASCII letters, digits, and hyphens.
short No One ASCII letter or digit without -.
env Yes Environment variable that supplies the option value.
default Yes Value used when neither the command line nor the environment supplies one.
values No Accepted values. Omit the field to allow any value. Include "..." to allow the listed values plus any other value. Without "...", the list is closed.
help Yes Non-empty text shown in suite help.

Option names, short names, and environment keys must be unique within a suite. An option environment key cannot also appear in path_replacements. Suite options cannot reuse the runner's global names: help (-h), query (-q), subgroup, iterations (-i), partitions (-n), batch-size (-s), mem-pool-type, memory-limit, sort-spill-reservation-bytes, debug (-d), simulate-latency, criterion, list, output (-o), save-baseline, path (-p), result-mode, or dry-run.

Benchmark configuration

benchmark_runner is the preferred interface for configuring and running SQL benchmarks. Run <suite> --help to see the common and suite-specific options:

cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- h2o --help

The runner maps suite options to the environment variables below for compatibility with benchmark files and direct Criterion runs. Direct cargo bench --bench sql invocations cannot accept custom arguments, so they still use environment variables.

DataFusion session settings are read through SessionConfig::from_env() in both CLI and direct Criterion runs. Use the standard DATAFUSION_* environment variables:

Setting Environment variable
Parallel partitions DATAFUSION_EXECUTION_TARGET_PARTITIONS
Batch size DATAFUSION_EXECUTION_BATCH_SIZE
Sort spill reservation (bytes) DATAFUSION_EXECUTION_SORT_SPILL_RESERVATION_BYTES
Prefer hash joins DATAFUSION_OPTIMIZER_PREFER_HASH_JOIN
Enable Piecewise Merge Join DATAFUSION_OPTIMIZER_ENABLE_PIECEWISE_MERGE_JOIN
Hash join buffering capacity DATAFUSION_EXECUTION_HASH_JOIN_BUFFERING_CAPACITY
Collect statistics DATAFUSION_EXECUTION_COLLECT_STATISTICS

Explicit CLI settings override environment settings; otherwise DataFusion's defaults apply. The old PARTITIONS, BATCH_SIZE, SORT_SPILL_RESERVATION_BYTES, PREFER_HASH_JOIN, and HASH_JOIN_BUFFERING_CAPACITY environment variables are no longer used. For example:

DATAFUSION_EXECUTION_TARGET_PARTITIONS=2 DATAFUSION_OPTIMIZER_PREFER_HASH_JOIN=false cargo bench --bench sql

Memory limits and pool types configure RuntimeEnv separately. MEMORY_LIMIT (or --memory-limit) takes precedence over DATAFUSION_RUNTIME_MEMORY_LIMIT.

The SQL benchmarking tool uses the following environment variables:

Environment Variable Description
BENCH_NAME The name of the benchmark suite to run. For example 'imdb'. This should correspond to a directory name in the sql_benchmarks directory.
BENCH_SUBGROUP The subgroup with the benchmark suite to run. For example 'window' to run the window subgroup of the h2o benchmark.
BENCH_QUERY A query number to run.
BENCH_PERSIST_RESULTS true/false to persist benchmark results. Results will be persisted in csv format so be cognizant of the size of the results.
BENCH_VALIDATE true/false to validate benchmark results against persisted results or result_query's. If both BENCH_PERSIST_RESULTS and BENCH_VALIDATE are true, persist mode runs and validation is skipped.
BENCH_RESULTS_FILE Criterion runs only: write a results JSON in the dfbench -o format to this path. Each query records its peak memory pool reservation (pool_peak_bytes, only under a memory limit, including one set with SET datafusion.runtime.memory_limit) and no timings, which stay with Criterion. bench.sh sets it to results/<name>/criterion/<file>.json.
DATA_DIR Root directory for benchmark data loaded by SQL benchmark files. When unset, uses data (relative to the benchmarks/ directory).
SIMULATE_LATENCY Simulate object store latency to mimic remote storage (e.g. S3). Adds random latency in the range 20-200ms to each object store operation.
MEM_POOL_TYPE The memory pool type to use, should be one of "fair" or "greedy".
MEMORY_LIMIT Memory limit (e.g. '100M', '1.5G'). If not specified, run all pre-defined memory limits for given query if there's any, otherwise run with no memory limit.

Example: run the H2O window benchmarks on the small CSV data files:

cargo run -p datafusion-benchmarks --release --bin benchmark_runner -- h2o --subgroup window --size small --format csv

Some benchmarks use custom environment variables as outlined below:

Name Description Default value
BENCH_SIZE Used in the tpch, sort-tpch and tpcds benchmarks. The size corresponds to the scale factor. 1
TPCH_FILE_TYPE Used in the tpch benchmark to specify which file type to query against. The valid options are csv, parquet and mem parquet
H2O_FILE_TYPE Used in the h2o benchmark to specify which file type to query against. The valid options are csv and parquet csv
CLICKBENCH_TYPE The type of partitioning for the clickbench benchmark. Valid options are single and partitioned single
H2O_BENCH_SIZE Used in the h2o benchmark. The valid options are small, medium and big small
BENCH_SORTED Used in the sort_tpch benchmark to indicate whether the lineitem table should be sorted. false
SORTED_BY Used in the clickbench_sorted benchmark to indicate the column to sort by. EventTime
SORTED_ORDER Used in the clickbench_sorted benchmark to indicate the sort order of the column. ASC
PRED_ROWS Used in the predicate_eval benchmark to size the synthetic table (the scale subgroup overrides this per query), and in the parquet_row_filter_skip benchmark to size the generated Parquet datasets (default 10000000 there). 1000000
PRED_FILL Used in the predicate_eval benchmark as the string-column width knob (filler chars per marker). 30
RG_SIZE Used in the parquet_row_filter_skip benchmark as the Parquet max_row_group_size for the generated datasets. 1000000
PSQ_ROWS Used in the projection_subquery benchmark to size the two inline tables. 30000
SPILL_VIEWS_LIMIT_REPEATED Used in the spill_views benchmark as the memory limit of the repeated subgroup. 40M
SPILL_VIEWS_LIMIT_DISTINCT Used in the spill_views benchmark as the memory limit of the distinct subgroup. 96M

How it works

The runner executes SQL benchmarks with its basic runner by default. Pass --criterion to gather statistics with Criterion.

Each individual benchmark is represented by a <name>.benchmark file that contains a number of directives instructing the tool on how to load data, run initializations, run assertions, run the benchmark, optionally persist and validate results, and finally run any cleanup if required.

Benchmark files support replacement variables in two forms:

  • string substitution with an optional default: ${ENV_VAR} and ${ENV_VAR:-default}.
  • if / else based on whether a replacement value is true or not (${ENV_VAR:-default|true value|false value}). In this form only the value true (case-insensitive) selects the true branch; any other supplied value selects the false branch. If the value is absent, the parser uses default to select the branch.

Comments in files are supported with lines starting with # or --.

Many if not most of the benchmarks are set up using templates to reduce duplication across the .benchmark files. For example here is one of the benchmark files for the h2o benchmark suite:

subgroup groupby

template sql_benchmarks/h2o/h2o.benchmark.template
QUERY_NUMBER=1
QUERY_NUMBER_PADDED=01

The template directive above defines the subgroup the benchmark is part of, sets two variables (QUERY_NUMBER and QUERY_NUMBER_PADDED) and points to a file containing more directives that are shared across the benchmark suite.

load sql_benchmarks/h2o/init/load_${BENCH_SUBGROUP:-groupby}_${BENCH_SIZE:-small}_${BENCH_FILE_TYPE:-csv}.sql

name Q${QUERY_NUMBER_PADDED}
group h2o

run sql_benchmarks/h2o/queries/${BENCH_SUBGROUP:-groupby}/q${QUERY_NUMBER_PADDED}.sql

result sql_benchmarks/h2o/results/${BENCH_SUBGROUP:-groupby}/${BENCH_SIZE:-small}/q${QUERY_NUMBER_PADDED}.csv

The above showcases the use of defaults for variables: ${NAME:-default}

Directives

DirectiveDescription
name

The name of the benchmark. This will be used as part of the display name used by criterion.

Example:

name Q${QUERY_NUMBER_PADDED}

The name directive also makes the value available to benchmark-file replacements as BENCH_NAME. This value is separate from the suite name passed to benchmark_runner.

group

The group name of the benchmark used for grouping benchmarks together.

Example:

group imdb
subgroup

The sub group name of the benchmark used for filtering to a specific sub group.

Example:

subgroup window
load

The load directive called during initialization of the benchmark. If a path to a file is provided on the same line as the load directive that path will be parsed and any sql statements in that file will be executed during initialization. If no path is specified the next line is required to be the sql statement to execute.

The load directive (including any following sql statement) must be followed by a blank line.

Example:

load sql_benchmarks/h2o/init/load_${BENCH_SUBGROUP:-groupby}_${BENCH_SIZE:-small}_${BENCH_FILE_TYPE:-csv}.sql
or
load
CREATE TABLE test AS (SELECT value as key FROM range(1000000) ORDER BY value);
init

The init directive is called after the load directive prior to benchmark execution. If a path to a file is provided on the same line as the init directive that path will be parsed and any sql statements in that file will be executed during the benchmark initialization. If no path is specified the next line is required to be the sql statement to execute.

The init directive (including any following sql statement) must be followed by a blank line.

Example:

init
set datafusion.execution.parquet.binary_as_string = true;
run

The run directive called during execution of the benchmark. If a path to a file is provided on the same line as the run directive that path will be parsed and any sql statements in that file will be executed during the benchmark run. If no path is specified the next line is required to be the sql statement to execute.

Multiple statements are allowed within a single run directive, however a benchmark file may contain only one run directive. When when persisting or validating results, only the last SELECT or WITH statement from that run directive will be used for comparison.

The run directive (including any following sql statement) must be followed by a blank line.

Example:

run sql_benchmarks/imdb/queries/${QUERY_NUMBER_PADDED}.sql
cleanup

The cleanup directive is called after all other directives and can be used to cleanup after the benchmark - e.g. to drop tables. If a path to a file is provided on the same line as the cleanup directive that path will be parsed and any sql statements in that file will be executed during cleanup. If no path is specified the next line is required to be the sql statement to execute.

The cleanup directive (including any following sql statement) must be followed by a blank line.

Example:

cleanup
DROP TABLE test;
expect_plan

The expect_plan directive will check the physical plan for the string provided on the same line. This can be used to validate that a particular join was used. The plan is rendered as EXPLAIN displays it, and the check runs once per benchmark, not once per iteration, so it adds no cost to the measured region.

Example:

expect_plan NestedLoopJoinExec
assert

The assert directive is run between the init and run directives and can be used to validate system state correctness prior to running the benchmark sql. The format is

assert II
SELECT name, value = 3 FROM information_schema.df_settings WHERE name IN ('datafusion.execution.target_partitions', 'datafusion.execution.planning_concurrency');
----
datafusion.execution.planning_concurrency true
datafusion.execution.target_partitions true

The number of I's corresponds to the number of columns in the result. The expected results can be either tab delimited or pipe delimited.

result_query

The result_query directive is run during the verify phase and can be used to verify a different set of results than any that might come from queries executed from the run directive. The format is the same as the assert directive above.

Example:

result_query III
SELECT COUNT(DISTINCT id2), SUM(r4), COUNT(*) FROM answer;
----
123 345 45

Note that the results of the run query are not automatically stored into a table in datafusion. If you want to verify a result from queries executed from the run directive those queries will have to be saved to a table directly using CREATE TABLE AS (..) or similar.

result

The result directive declares the expected result file used during verification. A path to a file is required on the same line as the result directive. The file is parsed only during verification, and must be a pipe-delimited CSV file with a header row. During verification, these expected rows are compared with the rows produced by the last saved SELECT or WITH statement from the run directive.

Example:

result sql_benchmarks/imdb/results/${QUERY_NUMBER_PADDED}.csv
template

The template directive allows for inclusion of another file in a benchmark file. A path to a file is required on the same line as the template directive which will be parsed as a benchmark file. Parameters can be passed to the template file using the format KEY=value, one per line after the template directive followed by a blank line.

Example:

template sql_benchmarks/smj/smj.benchmark.template
QUERY_NUMBER=1
QUERY_NUMBER_PADDED=01
include The include directive is similar to the template directive except that it does not support parameters.
echo

The echo directive allows for echoing a string to stdout during the execution of the benchmark and may be useful for debugging.

Example:

echo The value for batch size is ${DATAFUSION_EXECUTION_BATCH_SIZE:-8192}

Extending an existing benchmark suite

If you want to add a new query:

  • Create a new qXX.sql in the corresponding queries folder of the benchmark.
  • Add a new qXX.benchmark that references the appropriate template (clickbench.benchmark.template, h2o.benchmark.template, etc.).
  • (Optional) Add a new entry to the suite’s load script if the data set is different.
  • (Optional) Manually create a result csv to be compared against benchmark results during verification.

Adding a new benchmark suite

  • Create a new directory named for the new benchmark suite.
  • Within there create a <name>.benchmark for each individual benchmark.
  • Populate the benchmark with directives as described above. Use the other benchmarks as examples for standardization.
  • No rust files need to be updated to run the new benchmark suite.