Configuration properties

Generic Configuration

Property Name	Type	Default Value	Description
preferred_output_batch_bytes	integer	10MB	Preferred size of batches in bytes to be returned by operators from Operator::getOutput. It is used when an estimate of average row size is known. Otherwise preferred_output_batch_rows is used.
preferred_output_batch_rows	integer	1024	Preferred number of rows to be returned by operators from Operator::getOutput. It is used when an estimate of average row size is not known. When the estimate of average row size is known, preferred_output_batch_bytes is used.
max_output_batch_rows	integer	10000	Max number of rows that could be return by operators from Operator::getOutput. It is used when an estimate of average row size is known and preferred_output_batch_bytes is used to compute the number of output rows.
table_scan_getoutput_time_limit_ms	integer	5000	TableScan operator will exit getOutput() method after this many milliseconds even if it has no data to return yet. Zero means ‘no time limit’.
abandon_partial_aggregation_min_rows	integer	100,000	Number of input rows to receive before starting to check whether to abandon partial aggregation.
abandon_partial_aggregation_min_pct	integer	80	Abandons partial aggregation if number of groups equals or exceeds this percentage of the number of input rows.
abandon_partial_topn_row_number_min_rows	integer	100,000	Number of input rows to receive before starting to check whether to abandon partial TopNRowNumber.
abandon_partial_topn_row_number_min_pct	integer	80	Abandons partial TopNRowNumber if number of output rows equals or exceeds this percentage of the number of input rows.
session_timezone	string		User provided session timezone. Stores a string with the actual timezone name, e.g: “America/Los_Angeles”.
adjust_timestamp_to_session_timezone	bool	false	If true, timezone-less timestamp conversions (e.g. string to timestamp, when the string does not specify a timezone) will be adjusted to the user provided session_timezone (if any). For instance: if this option is true and user supplied “America/Los_Angeles”, then “1970-01-01” will be converted to -28800 instead of 0. Similarly, timestamp to date conversions will adhere to user ‘session_timezone’, e.g: Timestamp(0) to Date will be -1 (number of days since epoch) for “America/Los_Angeles”.
track_operator_cpu_usage	bool	true	Whether to track CPU usage for stages of individual operators. Can be expensive when processing small batches, e.g. < 10K rows.
hash_adaptivity_enabled	bool	true	If false, the ‘group by’ code is forced to use generic hash mode hashtable.
adaptive_filter_reordering_enabled	bool	true	If true, the conjunction expression can reorder inputs based on the time taken to calculate them.
max_local_exchange_buffer_size	integer	32MB	Used for backpressure to block local exchange producers when the local exchange buffer reaches or exceeds this size.
max_local_exchange_partition_count	integer	2^32	Limits the number of partitions created by a local exchange. Partitioning data too granularly can lead to poor performance. This setting allows increasing the task concurrency for all pipelines except the ones that require a local partitioning. Affects the number of drivers for pipelines containing LocalPartitionNode and cannot exceed the maximum number of pipeline drivers configured for the task.
exchange.max_buffer_size	integer	32MB	Size of buffer in the exchange client that holds data fetched from other nodes before it is processed. A larger buffer can increase network throughput for larger clusters and thus decrease query processing time at the expense of reducing the amount of memory available for other usage.
merge_exchange.max_buffer_size	integer	128MB	The aggregate buffer size (in bytes) across all exchange clients generated by the merge exchange operator, responsible for storing data retrieved from various nodes prior to processing. It is divided equally among all clients and has an upper and lower limit of 32MB and 1MB, respectively, per client. Enforced approximately, not strictly. A larger size can increase network throughput for larger clusters and thus decrease query processing time at the expense of reducing the amount of memory available for other usage.
max_page_partitioning_buffer_size	integer	32MB	The maximum size in bytes for the task’s buffered output when output is partitioned using hash of partitioning keys. See PartitionedOutputNode::Kind::kPartitioned. The producer Drivers are blocked when the buffered size exceeds this. The Drivers are resumed when the buffered size goes below OutputBufferManager::kContinuePct (90)% of this.
max_output_buffer_size	integer	32MB	The maximum size in bytes for the task’s buffered output. The producer Drivers are blocked when the buffered size exceeds this. The Drivers are resumed when the buffered size goes below OutputBufferManager::kContinuePct (90)% of this.
min_table_rows_for_parallel_join_build	integer	1000	The minimum number of table rows that can trigger the parallel hash join table build.
debug.validate_output_from_operators	bool	false	If set to true, then during execution of tasks, the output vectors of every operator are validated for consistency. This is an expensive check so should only be used for debugging. It can help debug issues where malformed vector cause failures or crashes by helping identify which operator is generating them.
enable_expression_evaluation_cache	bool	true	Whether to enable caches in expression evaluation. If set to true, optimizations including vector pools and evalWithMemo are enabled.
max_shared_subexpr_results_cached	integer	10	For a given shared subexpression, the maximum distinct sets of inputs we cache results for. Lambdas can call the same expression with different inputs many times, causing the results we cache to explode in size. Putting a limit contains the memory usage.
driver_cpu_time_slice_limit_ms	integer	0	If it is not zero, specifies the time limit that a driver can continuously run on a thread before yield. If it is zero, then it no limit.
prefixsort_normalized_key_max_bytes	integer	128	Maximum number of bytes to use for the normalized key in prefix-sort. Use 0 to disable prefix-sort.
prefixsort_min_rows	integer	130	Minimum number of rows to use prefix-sort. The default value has been derived using micro-benchmarking.

Expression Evaluation Configuration

Property Name	Type	Default Value	Description
expression.eval_simplified	boolean	false	Whether to use the simplified expression evaluation path.
expression.track_cpu_usage	boolean	false	Whether to track CPU usage for individual expressions (supported by call and cast expressions). Can be expensive when processing small batches, e.g. < 10K rows.
legacy_cast	bool	false	Enables legacy CAST semantics if set to true. CAST(timestamp AS varchar) uses ‘T’ as separator between date and time (instead of a space), and the year part is not padded.
cast_match_struct_by_name	bool	false	This flag makes the Row conversion to by applied in a way that the casting row field are matched by name instead of position.
expression.max_array_size_in_reduce	integer	100000	Reduce function will throw an error if encountered an array of size greater than this.
debug_disable_expression_with_peeling	bool	false	Disable optimization in expression evaluation to peel common dictionary layer from inputs. Should only be used for debugging.
debug_disable_common_sub_expressions	bool	false	Disable optimization in expression evaluation to re-use cached results for common sub-expressions. Should only be used for debugging.
debug_disable_expression_with_memoization	bool	false	Disable optimization in expression evaluation to re-use cached results between subsequent input batches that are dictionary encoded and have the same alphabet(underlying flat vector). Should only be used for debugging.
debug_disable_expression_with_lazy_inputs	bool	false	Disable optimization in expression evaluation to delay loading of lazy inputs unless required. Should only be used for debugging.

Memory Management

Property Name	Type	Default Value	Description
max_partial_aggregation_memory	integer	16MB	Maximum amount of memory in bytes for partial aggregation results. Increasing this value can result in less network transfer and lower CPU utilization by allowing more groups to be kept locally before being flushed, at the cost of additional memory usage.
max_extended_partial_aggregation_memory	integer	16MB	Maximum amount of memory in bytes for partial aggregation results if cardinality reduction is below partial_aggregation_reduction_ratio_threshold. Every time partial aggregate results size reaches max_partial_aggregation_memory bytes, the results are flushed. If cardinality reduction is below partial_aggregation_reduction_ratio_threshold, i.e. number of result rows / number of input rows > partial_aggregation_reduction_ratio_threshold, memory limit for partial aggregation is automatically doubled up to max_extended_partial_aggregation_memory. This adaptation is disabled by default, since the value of max_extended_partial_aggregation_memory equals the value of max_partial_aggregation_memory. Specify higher value for max_extended_partial_aggregation_memory to enable.

Spilling

Property Name	Type	Default Value	Description
spill_enabled	boolean	false	Spill memory to disk to avoid exceeding memory limits for the query.
aggregation_spill_enabled	boolean	true	When spill_enabled is true, determines whether HashAggregation operator can spill to disk under memory pressure.
join_spill_enabled	boolean	true	When spill_enabled is true, determines whether HashBuild and HashProbe operators can spill to disk under memory pressure.
order_by_spill_enabled	boolean	true	When spill_enabled is true, determines whether OrderBy operator can spill to disk under memory pressure.
window_spill_enabled	boolean	true	When spill_enabled is true, determines whether Window operator can spill to disk under memory pressure.
row_number_spill_enabled	boolean	true	When spill_enabled is true, determines whether RowNumber operator can spill to disk under memory pressure.
topn_row_number_spill_enabled	boolean	true	When spill_enabled is true, determines whether TopNRowNumber operator can spill to disk under memory pressure.
writer_spill_enabled	boolean	true	When writer_spill_enabled is true, determines whether TableWriter operator can flush the buffered data to disk under memory pressure.
aggregation_spill_memory_threshold	integer	0	Maximum amount of memory in bytes that a final aggregation can use before spilling. 0 means unlimited.
join_spill_memory_threshold	integer	0	Maximum amount of memory in bytes that a hash join build side can use before spilling. 0 means unlimited.
order_by_spill_memory_threshold	integer	0	Maximum amount of memory in bytes that an order by can use before spilling. 0 means unlimited.
writer_flush_threshold_bytes	integer	96MB	Minimum memory footprint size required to reclaim memory from a file writer by flushing its buffered data to disk.
min_spillable_reservation_pct	integer	5	The minimal available spillable memory reservation in percentage of the current memory usage. Suppose the current memory usage size of M, available memory reservation size of N and min reservation percentage of P, if M * P / 100 > N, then spiller operator needs to grow the memory reservation with percentage of ‘spillable_reservation_growth_pct’ (see below). This ensures we have sufficient amount of memory reservation to process the large input outlier.
spillable_reservation_growth_pct	integer	10	The spillable memory reservation growth percentage of the current memory usage. Suppose a growth percentage of N and the current memory usage size of M, the next memory reservation size will be M * (1 + N / 100). After growing the memory reservation K times, the memory reservation size will be M * (1 + N / 100) ^ K. Hence the memory reservation grows along a series of powers of (1 + N / 100). If the memory reservation fails, it starts spilling.
max_spill_level	integer	1	The maximum allowed spilling level with zero being the initial spilling level. Applies to hash join build spilling which might use recursive spilling when the build table is very large. -1 means unlimited. In this case an extremely large query might run out of spilling partition bits. The max spill level can be used to prevent a query from using too much io and cpu resources.
max_spill_run_rows	integer	12582912	The max number of rows to fill and spill for each spill run. This is used to cap the memory used for spilling. If it is zero, then there is no limit and spilling might run out of memory. Based on offline test results, the default value is set to 12 million rows which uses ~128MB memory when to fill a spill run. Relation between spill rows and memory usage are as follows: 12 million rows: 128 MB 30 million rows: 256 MB 60 million rows: 512 MB
max_spill_file_size	integer	0	The maximum allowed spill file size. Zero means unlimited.
max_spill_bytes	integer	107374182400	The max spill bytes limit set for each query. This is used to cap the storage used for spilling. If it is zero, then there is no limit and spilling might exhaust the storage or takes too long to run. The default value is set to 100 GB.
spill_write_buffer_size	integer	4MB	The maximum size in bytes to buffer the serialized spill data before write to disk for IO efficiency. If set to zero, buffering is disabled.
spill_read_buffer_size	integer	1MB	The buffer size in bytes to read from one spilled file. If the underlying filesystem supports async read, we do read-ahead with double buffering, which doubles the buffer used to read from each spill file.
min_spill_run_size	integer	256MB	The minimum spill run size (bytes) limit used to select partitions for spilling. The spiller tries to spill a previously spilled partitions if its data size exceeds this limit, otherwise it spills the partition with most data. If the limit is zero, then the spiller always spills a previously spilled partition if it has any data. This is to avoid spill from a partition with a small amount of data which might result in generating too many small spilled files.
spill_compression_codec	string	none	Specifies the compression algorithm type to compress the spilled data before write to disk to trade CPU for IO efficiency. The supported compression codecs are: ZLIB, SNAPPY, LZO, ZSTD, LZ4 and GZIP. NONE means no compression.
spill_prefixsort_enabled	bool	false	Enable the prefix sort or fallback to timsort in spill. The prefix sort is faster than std::sort but requires the memory to build normalized prefix keys, which might have potential risk of running out of server memory.
spiller_start_partition_bit	integer	29	The start partition bit which is used with spiller_partition_bits together to calculate the spilling partition number.
spiller_num_partition_bits	integer	3	The number of bits (N) used to calculate the spilling partition number for hash join and RowNumber: 2 ^ N. At the moment the maximum value is 3, meaning we only support up to 8-way spill partitioning.ing.
testing.spill_pct	integer	0	Percentage of aggregation or join input batches that will be forced to spill for testing. 0 means no extra spilling.

Table Scan

Property Name	Type	Default Value	Description
max_split_preload_per_driver	integer	2	Maximum number of splits to preload per driver. Set to 0 to disable preloading.

Table Writer

Property Name	Type	Default Value	Description
task_writer_count	integer	1	The number of parallel table writer threads per task.
task_partitioned_writer_count	integer	task_writer_count	The number of parallel table writer threads per task for partitioned table writes. If not set, use ‘task_writer_count’ as default.

Hive Connector

Hive Connector config is initialized on velox runtime startup and is shared among queries as the default config. Each query can override the config by setting corresponding query session properties such as in Prestissimo.

Configuration Property Name	Session Property Name	Type	Default Value	Description
hive.max-partitions-per-writers		integer	100	Maximum number of (bucketed) partitions per a single table writer instance.
insert-existing-partitions-behavior	insert_existing_partitions_behavior	string	ERROR	Allowed values: OVERWRITE, ERROR. The behavior on insert existing partitions. This property only derives the update mode field of the table writer operator output. OVERWRITE sets the update mode to indicate overwriting a partition if exists. ERROR sets the update mode to indicate error throwing if writing to an existing partition.
hive.immutable-partitions		bool	false	True if appending data to an existing unpartitioned table is allowed. Currently this configuration does not support appending to existing partitions.
file-column-names-read-as-lower-case		bool	false	True if reading the source file column names as lower case, and planner should guarantee the input column name and filter is also lower case to achive case-insensitive read.
partition_path_as_lower_case		bool	true	If true, the partition directory will be converted to lowercase when executing a table write operation.
allow-null-partition-keys	allow_null_partition_keys	bool	true	Determines whether null values for partition keys are allowed or not. If not, fails with “Partition key must not be null” error message when writing data with null partition key. Null check for partitioning key should be used only when partitions are generated dynamically during query execution. For queries that write to fixed partitions, this check should happen much earlier before the Velox execution even starts.
ignore_missing_files		bool	false	If true, splits that refer to missing files don’t generate errors and are processed as empty splits.
max-coalesced-bytes		integer	128MB	Maximum size in bytes to coalesce requests to be fetched in a single request.
max-coalesced-distance-bytes		integer	512KB	Maximum distance in bytes between chunks to be fetched that may be coalesced into a single request.
load-quantum		integer	8MB	Define the size of each coalesce load request. E.g. in Parquet scan, if it’s bigger than rowgroup size then the whole row group can be fetched together. Otherwise, the row group will be fetched column chunk by column chunk
num-cached-file-handles		integer	20000	Maximum number of entries in the file handle cache. The value must be non-negative. Zero value indicates infinite cache capacity.
file-handle-cache-enabled		bool	true	Enables caching of file handles if true. Disables caching if false. File handle cache should be disabled if files are not immutable, i.e. file content may change while file path stays the same.
sort-writer-max-output-rows	sort_writer_max_output_rows	integer	1024	Maximum number of rows for sort writer in one batch of output. This is to limit the memory usage of sort writer.
sort-writer-max-output-bytes	sort_writer_max_output_bytes	string	10MB	Maximum bytes for sort writer in one batch of output. This is to limit the memory usage of sort writer.
file-preload-threshold		integer	8MB	Usually Velox fetches the meta data firstly then fetch the rest of file. But if the file is very small, Velox can fetch the whole file directly to avoid multiple IO requests. The parameter controls the threshold when whole file is fetched.
footer-estimated-size		integer	1MB	Define the estimation of footer size in ORC and Parquet format. The footer data includes version, schema, and meta data for every columns which may or may not need to be fetched later. The parameter controls the size when footer is fetched each time. Bigger value can decrease the IO requests but may fetch more useless meta data.
hive.orc.writer.stripe-max-size	orc_optimized_writer_max_stripe_size	string	64M	Maximum stripe size in orc writer.
hive.orc.writer.dictionary-max-memory	orc_optimized_writer_max_dictionary_memory	string	16M	Maximum dictionary memory that can be used in orc writer.
hive.orc.writer.integer-dictionary-encoding-enabled	orc_optimized_writer_integer_dictionary_encoding_enabled	bool	true	Whether or not dictionary encoding of integer types should be used by the ORC writer.
hive.orc.writer.string-dictionary-encoding-enabled	orc_optimized_writer_string_dictionary_encoding_enabled	bool	true	Whether or not dictionary encoding of string types should be used by the ORC writer.
hive.parquet.writer.timestamp-unit	hive.parquet.writer.timestamp_unit	tinyint	9	Timestamp unit used when writing timestamps into Parquet through Arrow bridge. Valid values are 0 (second), 3 (millisecond), 6 (microsecond), 9 (nanosecond).
hive.orc.writer.linear-stripe-size-heuristics	orc_writer_linear_stripe_size_heuristics	bool	true	Enables historical based stripe size estimation after compression.
hive.orc.writer.min-compression-size	orc_writer_min_compression_size	integer	1024	Minimal number of items in an encoded stream.
hive.orc.writer.compression-level	orc_optimized_writer_compression_level	tinyint	3 for ZSTD and 4 for ZLIB	The compression level to use with ZLIB and ZSTD.
cache.no_retention	cache.no_retention	bool	false	If true, evict out a query scanned data out of in-memory cache right after the access, and also skip staging to the ssd cache. This helps to prevent the cache space pollution from the one-time table scan by large batch query when mixed running with interactive query which has high data locality.

Amazon S3 Configuration

Property Name	Type	Default Value	Description
hive.s3.use-instance-credentials	bool	false	Use the EC2 metadata service to retrieve API credentials. This works with IAM roles in EC2.
hive.s3.aws-access-key	string		Default AWS access key to use.
hive.s3.aws-secret-key	string		Default AWS secret key to use.
hive.s3.endpoint	string	us-east-1	The S3 storage endpoint server. This can be used to connect to an S3-compatible storage system instead of AWS.
hive.s3.path-style-access	bool	false	Use path-style access for all requests to the S3-compatible storage. This is for S3-compatible storage that doesn’t support virtual-hosted-style access.
hive.s3.ssl.enabled	bool	true	Use HTTPS to communicate with the S3 API.
hive.s3.log-level	string	FATAL	Allowed values: “OFF”, “FATAL”, “ERROR”, “WARN”, “INFO”, “DEBUG”, “TRACE” Granularity of logging generated by the AWS C++ SDK library.
hive.s3.iam-role	string		IAM role to assume.
hive.s3.iam-role-session-name	string	velox-session	Session name associated with the IAM role.
hive.s3.use-proxy-from-env	bool	false	Utilize the configuration of the environment variables http_proxy, https_proxy, and no_proxy for use with the S3 API.
hive.s3.connect-timeout	string		Socket connect timeout.
hive.s3.socket-timeout	string		Socket read timeout.
hive.s3.max-connections	integer		Maximum concurrent TCP connections for a single http client.
hive.s3.max-attempts	integer		Maximum attempts for connections to a single http client, work together with retry-mode. By default, it’s 3 for standard/adaptive mode and 10 for legacy mode.
hive.s3.retry-mode	string		Allowed values: “standard”, “adaptive”, “legacy”. By default it’s empty, S3 client will be created with RetryStrategy. Legacy mode only enables throttled retry for transient errors. Standard mode is built on top of legacy mode and has throttled retry enabled for throttling errors apart from transient errors. Adaptive retry mode dynamically limits the rate of AWS requests to maximize success rate.

Bucket Level Configuration

All “hive.s3.*” config (except “hive.s3.log-level”) can be set on a per-bucket basis. The bucket-specific option is set by replacing the “hive.s3.” prefix on a config with “hive.s3.bucket.BUCKETNAME.”, where BUCKETNAME is the name of the bucket. e.g. the endpoint for a bucket named “velox” can be specified by the config “hive.s3.bucket.velox.endpoint”. When connecting to a bucket, all options explicitly set will override the base “hive.s3.” values. These semantics are similar to the Apache Hadoop-Aws module.

Google Cloud Storage Configuration

Property Name	Type	Default Value	Description
hive.gcs.endpoint	string		The GCS storage URI.
hive.gcs.json-key-file-path	string		The GCS service account configuration JSON key file.
hive.gcs.max-retry-count	integer		The GCS maximum retry counter of transient errors.
hive.gcs.max-retry-time	string		The GCS maximum time allowed to retry transient errors.

Azure Blob Storage Configuration

Property Name	Type	Default Value	Description
fs.azure.account.key.<storage-account>.dfs.core.windows.net	string		The credentials to access the specific Azure Blob Storage account, replace <storage-account> with the name of your Azure Storage account. This property aligns with how Spark configures Azure account key credentials for accessing Azure storage, by setting this property multiple times with different storage account names, you can access multiple Azure storage accounts.

Presto-specific Configuration

Property Name	Type	Default Value	Description
presto.array_agg.ignore_nulls	bool	false	If true, array_agg function ignores null inputs.

Spark-specific Configuration

Property Name	Type	Default Value	Description
spark.legacy_size_of_null	bool	true	If false, size function returns null for null input.
spark.bloom_filter.expected_num_items	integer	1000000	The default number of expected items for the bloom filter in bloom_filter_agg() function.
spark.bloom_filter.num_bits	integer	8388608	The default number of bits to use for the bloom filter in bloom_filter_agg() function.
spark.bloom_filter.max_num_bits	integer	4194304	The maximum number of bits to use for the bloom filter in bloom_filter_agg() function, the value of this config can not exceed the default value.
spark.partition_id	integer		The current task’s Spark partition ID. It’s set by the query engine (Spark) prior to task execution.
spark.legacy_date_formatter	bool	false	If true, Simple Date Format is used for time formatting and parsing. Joda date formatter is used by default. Joda date formatter performs strict checking of its input and uses different pattern string. For example, the 2015-07-22 10:00:00 timestamp cannot be parsed if pattern is yyyy-MM-dd because the parser does not consume whole input. Another example is that the ‘W’ pattern, which means week in month, is not supported. For more differences, see #10354.

Tracing

Property Name	Type	Default Value	Description
query_trace_enabled	bool	true	If true, enable query tracing.
query_trace_dir	string		The root directory to store the tracing data and metadata for a query.
query_trace_node_ids	string		A comma-separated list of plan node ids whose input data will be trace. If it is empty, then we only trace the query metadata which includes the query plan and configs etc.
query_trace_task_reg_exp	string		The regexp of traced task id. We only enable trace on a task if its id matches.
query_trace_max_bytes	integer	0	The max trace bytes limit. Tracing is disabled if zero.