Delta Lake connector#
The Delta Lake connector allows querying data stored in the Delta Lake format, including Databricks Delta Lake. The connector can natively read the Delta Lake transaction log and thus detect when external systems change data.
Requirements#
To connect to Databricks Delta Lake, you need:
Tables written by Databricks Runtime 7.3 LTS, 9.1 LTS, 10.4 LTS, 11.3 LTS, 12.2 LTS and 13.3 LTS are supported.
Deployments using AWS, HDFS, Azure Storage, and Google Cloud Storage (GCS) are fully supported.
Network access from the coordinator and workers to the Delta Lake storage.
Access to the Hive metastore service (HMS) of Delta Lake or a separate HMS, or a Glue metastore.
Network access to the HMS from the coordinator and workers. Port 9083 is the default port for the Thrift protocol used by the HMS.
Data files stored in the Parquet file format. These can be configured using file format configuration properties per catalog.
General configuration#
To configure the Delta Lake connector, create a catalog properties file
etc/catalog/example.properties that references the delta_lake
connector and defines a metastore. You must configure a metastore for table
metadata. If you are using a Hive metastore,
hive.metastore.uri must be configured:
connector.name=delta_lake
hive.metastore.uri=thrift://example.net:9083
If you are using AWS Glue as your metastore, you
must instead set hive.metastore to glue:
connector.name=delta_lake
hive.metastore=glue
Each metastore type has specific configuration properties along with general metastore configuration properties.
The connector recognizes Delta Lake tables created in the metastore by the Databricks runtime. If non-Delta Lake tables are present in the metastore as well, they are not visible to the connector.
To configure access to S3 and S3-compatible storage, Azure storage, and others, consult the appropriate section of the Hive documentation:
Delta Lake general configuration properties#
The following configuration properties are all using reasonable, tested default values. Typical usage does not require you to configure them.
Property name |
Description |
Default |
|---|---|---|
|
Frequency of checks for metadata updates equivalent to transactions to update the metadata cache specified in duration. |
|
|
The maximum number of Delta table metadata entries to cache. |
|
|
Amount of memory allocated for caching information about files. Must
be specified in data size values such as |
|
|
Caching duration for active files that correspond to the Delta Lake tables. |
|
|
The compression codec to be used when writing new data files. Possible values are:
The equivalent catalog session property is |
|
|
Maximum number of partitions per writer. |
|
|
Hide information about tables that are not managed by Delta Lake. Hiding only applies to tables with the metadata managed in a Glue catalog, and does not apply to usage with a Hive metastore service. |
|
|
Enable write support for all supported file systems. Specifically, take note of the warning about concurrency and checkpoints. |
|
|
Default integer count to write transaction log checkpoint entries. If
the value is set to N, then checkpoints are written after every Nth
statement performing table writes. The value can be overridden for a
specific table with the |
|
|
Name of the catalog to which |
|
|
Enable writing row statistics to checkpoint files. |
|
|
Duration to wait for completion of dynamic filtering during split generation.
The equivalent catalog session property is
|
|
|
Enables Table statistics for
performance improvements. The equivalent catalog session property
is |
|
|
Enable statistics collection with ANALYZE and
use of extended statistics. The equivalent catalog session property
is |
|
|
Enable collection of extended statistics for write operations.
The equivalent catalog session property is
|
|
|
Maximum number of metastore data objects per transaction in the Hive metastore cache. |
|
|
Whether schema locations are deleted when Trino can’t determine whether they contain external files. |
|
|
Time zone for Parquet read and write. |
JVM default |
|
Target maximum size of written files; the actual size could be larger.
The equivalent catalog session property is |
|
|
Use randomized, unique table locations. |
|
|
Enable to allow users to call the |
|
|
Minimum retention threshold for the files taken into account
for removal by the VACUUM procedure.
The equivalent catalog session property is
|
|
Catalog session properties#
The following table describes catalog session properties supported by the Delta Lake connector:
Property name |
Description |
Default |
|---|---|---|
|
The maximum block size used when reading Parquet files. |
|
|
The maximum block size created by the Parquet writer. |
|
|
The maximum page size created by the Parquet writer. |
|
|
Maximum number of rows processed by the Parquet writer in a batch. |
|
|
Read only projected fields from row columns while performing |
|
Type mapping#
Because Trino and Delta Lake each support types that the other does not, this connector modifies some types when reading or writing data. Data types might not map the same way in both directions between Trino and the data source. Refer to the following sections for type mapping in each direction.
See the Delta Transaction Log specification for more information about supported data types in the Delta Lake table format specification.
Delta Lake to Trino type mapping#
The connector maps Delta Lake types to the corresponding Trino types following this table:
Delta Lake type |
Trino type |
|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
No other types are supported.
Trino to Delta Lake type mapping#
The connector maps Trino types to the corresponding Delta Lake types following this table:
Trino type |
Delta Lake type |
|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
No other types are supported.
Security#
The Delta Lake connector allows you to choose one of several means of providing authorization at the catalog level. You can select a different type of authorization check in different Delta Lake catalog files.
SQL support#
The connector provides read and write access to data and metadata in Delta Lake. In addition to the globally available and read operation statements, the connector supports the following features:
-
Data management, see details for Delta Lake data management
Schema and table management, see details for Delta Lake schema and table management
Procedures#
Use the CALL statement to perform data manipulation or
administrative tasks. Procedures are available in the system schema of each
catalog. The following code snippet displays how to call the
example_procedure in the examplecatalog catalog:
CALL examplecatalog.system.example_procedure()
Register table#
The connector can register table into the metastore with existing transaction logs and data files.
The system.register_table procedure allows the caller to register an
existing Delta Lake table in the metastore, using its existing transaction logs
and data files:
CALL example.system.register_table(schema_name => 'testdb', table_name => 'customer_orders', table_location => 's3://my-bucket/a/path')
To prevent unauthorized users from accessing data, this procedure is disabled by
default. The procedure is enabled only when
delta.register-table-procedure.enabled is set to true.
Unregister table#
The connector can unregister existing Delta Lake tables from the metastore.
The procedure system.unregister_table allows the caller to unregister an
existing Delta Lake table from the metastores without deleting the data:
CALL example.system.unregister_table(schema_name => 'testdb', table_name => 'customer_orders')
Flush metadata cache#
system.flush_metadata_cache()Flushes all metadata caches.
system.flush_metadata_cache(schema_name => ..., table_name => ...)Flushes metadata cache entries of a specific table. Procedure requires passing named parameters.
VACUUM#
The VACUUM procedure removes all old files that are not in the transaction
log, as well as files that are not needed to read table snapshots newer than the
current time minus the retention period defined by the retention period
parameter.
Users with INSERT and DELETE permissions on a table can run VACUUM
as follows:
CALL example.system.vacuum('exampleschemaname', 'exampletablename', '7d');
All parameters are required and must be presented in the following order:
Schema name
Table name
Retention period
The delta.vacuum.min-retention configuration property provides a safety
measure to ensure that files are retained as expected. The minimum value for
this property is 0s. There is a minimum retention session property as well,
vacuum_min_retention.
Data management#
You can use the connector to INSERT, DELETE, UPDATE, and MERGE data in Delta Lake tables.
Write operations are supported for tables stored on the following systems:
Azure ADLS Gen2, Google Cloud Storage
Writes to the Azure ADLS Gen2 and Google Cloud Storage are enabled by default. Trino detects write collisions on these storage systems when writing from multiple Trino clusters, or from other query engines.
S3 and S3-compatible storage
Writes to Amazon S3 and S3-compatible storage must be enabled with the
delta.enable-non-concurrent-writesproperty. Writes to S3 can safely be made from multiple Trino clusters; however, write collisions are not detected when writing concurrently from other Delta Lake engines. You must make sure that no concurrent data modifications are run to avoid data corruption.
Schema and table management#
The Schema and table management functionality includes support for:
ALTER TABLE, see details for Delta Lake ALTER TABLE
The connector supports creating schemas. You can create a schema with or without a specified location.
You can create a schema with the CREATE SCHEMA statement and the
location schema property. Tables in this schema are located in a
subdirectory under the schema location. Data files for tables in this schema
using the default location are cleaned up if the table is dropped:
CREATE SCHEMA example.example_schema
WITH (location = 's3://my-bucket/a/path');
Optionally, the location can be omitted. Tables in this schema must have a location included when you create them. The data files for these tables are not removed if the table is dropped:
CREATE SCHEMA example.example_schema;
When Delta Lake tables exist in storage but not in the metastore, Trino can be used to register the tables:
CREATE TABLE example.default.example_table (
dummy BIGINT
)
WITH (
location = '...'
)
Columns listed in the DDL, such as dummy in the preceding example, are
ignored. The table schema is read from the transaction log instead. If the
schema is changed by an external system, Trino automatically uses the new
schema.
Warning
Using CREATE TABLE with an existing table content is deprecated, instead
use the system.register_table procedure. The CREATE TABLE ... WITH (location=...) syntax can be temporarily re-enabled using the
delta.legacy-create-table-with-existing-location.enabled catalog
configuration property or
legacy_create_table_with_existing_location_enabled catalog session
property.
If the specified location does not already contain a Delta table, the connector automatically writes the initial transaction log entries and registers the table in the metastore. As a result, any Databricks engine can write to the table:
CREATE TABLE example.default.new_table (id BIGINT, address VARCHAR);
The Delta Lake connector also supports creating tables using the CREATE TABLE AS syntax.
The connector supports the following ALTER TABLE statements.
ALTER TABLE EXECUTE#
The connector supports the following commands for use with ALTER TABLE EXECUTE.
optimize#
The optimize command is used for rewriting the content of the specified
table so that it is merged into fewer but larger files. If the table is
partitioned, the data compaction acts separately on each partition selected for
optimization. This operation improves read performance.
All files with a size below the optional file_size_threshold parameter
(default value for the threshold is 100MB) are merged:
ALTER TABLE test_table EXECUTE optimize
The following statement merges files in a table that are under 10 megabytes in size:
ALTER TABLE test_table EXECUTE optimize(file_size_threshold => '10MB')
You can use a WHERE clause with the columns used to partition the table
to filter which partitions are optimized:
ALTER TABLE test_partitioned_table EXECUTE optimize
WHERE partition_key = 1
ALTER TABLE RENAME TO#
The connector only supports the ALTER TABLE RENAME TO statement when met with
one of the following conditions:
The table type is external.
The table is backed by a metastore that does not perform object storage operations, for example, AWS Glue or Thrift.
Table properties#
The following table properties are available for use:
Property name |
Description |
|---|---|
|
File system location URI for the table. |
|
Set partition columns. |
|
Set the checkpoint interval in seconds. |
|
Enables storing change data feed entries. |
|
Column mapping mode. Possible values are:
Defaults to |
The following example uses all available table properties:
CREATE TABLE example.default.example_partitioned_table
WITH (
location = 's3://my-bucket/a/path',
partitioned_by = ARRAY['regionkey'],
checkpoint_interval = 5,
change_data_feed_enabled = false,
column_mapping_mode = 'name'
)
AS SELECT name, comment, regionkey FROM tpch.tiny.nation;
Metadata tables#
The connector exposes several metadata tables for each Delta Lake table. These metadata tables contain information about the internal structure of the Delta Lake table. You can query each metadata table by appending the metadata table name to the table name:
SELECT * FROM "test_table$history"
$history table#
The $history table provides a log of the metadata changes performed on
the Delta Lake table.
You can retrieve the changelog of the Delta Lake table test_table
by using the following query:
SELECT * FROM "test_table$history"
version | timestamp | user_id | user_name | operation | operation_parameters | cluster_id | read_version | isolation_level | is_blind_append
---------+---------------------------------------+---------+-----------+--------------+---------------------------------------+---------------------------------+--------------+-------------------+----------------
2 | 2023-01-19 07:40:54.684 Europe/Vienna | trino | trino | WRITE | {queryId=20230119_064054_00008_4vq5t} | trino-406-trino-coordinator | 2 | WriteSerializable | true
1 | 2023-01-19 07:40:41.373 Europe/Vienna | trino | trino | ADD COLUMNS | {queryId=20230119_064041_00007_4vq5t} | trino-406-trino-coordinator | 0 | WriteSerializable | true
0 | 2023-01-19 07:40:10.497 Europe/Vienna | trino | trino | CREATE TABLE | {queryId=20230119_064010_00005_4vq5t} | trino-406-trino-coordinator | 0 | WriteSerializable | true
The output of the query has the following history columns:
Name |
Type |
Description |
|---|---|---|
|
|
The version of the table corresponding to the operation |
|
|
The time when the table version became active |
|
|
The identifier for the user which performed the operation |
|
|
The username for the user which performed the operation |
|
|
The name of the operation performed on the table |
|
|
Parameters of the operation |
|
|
The ID of the cluster which ran the operation |
|
|
The version of the table which was read in order to perform the operation |
|
|
The level of isolation used to perform the operation |
|
|
Whether or not the operation appended data |
$properties table#
The $properties table provides access to Delta Lake table configuration,
table features and table properties. The table rows are key/value pairs.
You can retrieve the properties of the Delta
table test_table by using the following query:
SELECT * FROM "test_table$properties"
key | value |
----------------------------+-----------------+
delta.minReaderVersion | 1 |
delta.minWriterVersion | 4 |
delta.columnMapping.mode | name |
delta.feature.columnMapping | supported |
Metadata columns#
In addition to the defined columns, the Delta Lake connector automatically exposes metadata in a number of hidden columns in each table. You can use these columns in your SQL statements like any other column, e.g., they can be selected directly or used in conditional statements.
$pathFull file system path name of the file for this row.
$file_modified_timeDate and time of the last modification of the file for this row.
$file_sizeSize of the file for this row.
Fault-tolerant execution support#
The connector supports Fault-tolerant execution of query processing. Read and write operations are both supported with any retry policy.
Table functions#
The connector provides the following table functions:
table_changes#
Allows reading Change Data Feed (CDF) entries to expose row-level changes
between two versions of a Delta Lake table. When the change_data_feed_enabled
table property is set to true on a specific Delta Lake table,
the connector records change events for all data changes on the table.
This is how these changes can be read:
SELECT
*
FROM
TABLE(
system.table_changes(
schema_name => 'test_schema',
table_name => 'tableName',
since_version => 0
)
);
schema_name - type VARCHAR, required, name of the schema for which the function is called
table_name - type VARCHAR, required, name of the table for which the function is called
since_version - type BIGINT, optional, version from which changes are shown, exclusive
In addition to returning the columns present in the table, the function returns the following values for each change event:
_change_typeGives the type of change that occurred. Possible values are
insert,delete,update_preimageandupdate_postimage.
_commit_versionShows the table version for which the change occurred.
_commit_timestampRepresents the timestamp for the commit in which the specified change happened.
This is how it would be normally used:
Create table:
CREATE TABLE test_schema.pages (page_url VARCHAR, domain VARCHAR, views INTEGER)
WITH (change_data_feed_enabled = true);
Insert data:
INSERT INTO test_schema.pages
VALUES
('url1', 'domain1', 1),
('url2', 'domain2', 2),
('url3', 'domain1', 3);
INSERT INTO test_schema.pages
VALUES
('url4', 'domain1', 400),
('url5', 'domain2', 500),
('url6', 'domain3', 2);
Update data:
UPDATE test_schema.pages
SET domain = 'domain4'
WHERE views = 2;
Select changes:
SELECT
*
FROM
TABLE(
system.table_changes(
schema_name => 'test_schema',
table_name => 'pages',
since_version => 1
)
)
ORDER BY _commit_version ASC;
The preceding sequence of SQL statements returns the following result:
page_url | domain | views | _change_type | _commit_version | _commit_timestamp
url4 | domain1 | 400 | insert | 2 | 2023-03-10T21:22:23.000+0000
url5 | domain2 | 500 | insert | 2 | 2023-03-10T21:22:23.000+0000
url6 | domain3 | 2 | insert | 2 | 2023-03-10T21:22:23.000+0000
url2 | domain2 | 2 | update_preimage | 3 | 2023-03-10T22:23:24.000+0000
url2 | domain4 | 2 | update_postimage | 3 | 2023-03-10T22:23:24.000+0000
url6 | domain3 | 2 | update_preimage | 3 | 2023-03-10T22:23:24.000+0000
url6 | domain4 | 2 | update_postimage | 3 | 2023-03-10T22:23:24.000+0000
The output shows what changes happen in which version.
For example in version 3 two rows were modified, first one changed from
('url2', 'domain2', 2) into ('url2', 'domain4', 2) and the second from
('url6', 'domain2', 2) into ('url6', 'domain4', 2).
If since_version is not provided the function produces change events
starting from when the table was created.
SELECT
*
FROM
TABLE(
system.table_changes(
schema_name => 'test_schema',
table_name => 'pages'
)
)
ORDER BY _commit_version ASC;
The preceding SQL statement returns the following result:
page_url | domain | views | _change_type | _commit_version | _commit_timestamp
url1 | domain1 | 1 | insert | 1 | 2023-03-10T20:21:22.000+0000
url2 | domain2 | 2 | insert | 1 | 2023-03-10T20:21:22.000+0000
url3 | domain1 | 3 | insert | 1 | 2023-03-10T20:21:22.000+0000
url4 | domain1 | 400 | insert | 2 | 2023-03-10T21:22:23.000+0000
url5 | domain2 | 500 | insert | 2 | 2023-03-10T21:22:23.000+0000
url6 | domain3 | 2 | insert | 2 | 2023-03-10T21:22:23.000+0000
url2 | domain2 | 2 | update_preimage | 3 | 2023-03-10T22:23:24.000+0000
url2 | domain4 | 2 | update_postimage | 3 | 2023-03-10T22:23:24.000+0000
url6 | domain3 | 2 | update_preimage | 3 | 2023-03-10T22:23:24.000+0000
url6 | domain4 | 2 | update_postimage | 3 | 2023-03-10T22:23:24.000+0000
You can see changes that occurred at version 1 as three inserts. They are
not visible in the previous statement when since_version value was set to 1.
Performance#
The connector includes a number of performance improvements detailed in the following sections:
Support for write partitioning.
Table statistics#
Use ANALYZE statements in Trino to populate data size and number of distinct values (NDV) extended table statistics in Delta Lake. The minimum value, maximum value, value count, and null value count statistics are computed on the fly out of the transaction log of the Delta Lake table. The cost-based optimizer then uses these statistics to improve query performance.
Extended statistics enable a broader set of optimizations, including join
reordering. The controlling catalog property delta.table-statistics-enabled
is enabled by default. The equivalent catalog session property is statistics_enabled.
Each ANALYZE statement updates the table statistics incrementally, so only
the data changed since the last ANALYZE is counted. The table statistics are
not automatically updated by write operations such as INSERT, UPDATE,
and DELETE. You must manually run ANALYZE again to update the table
statistics.
To collect statistics for a table, execute the following statement:
ANALYZE table_schema.table_name;
To recalculate from scratch the statistics for the table use additional parameter mode:
ANALYZE table_schema.table_name WITH(mode = ‘full_refresh’);
There are two modes available full_refresh and incremental.
The procedure use incremental by default.
To gain the most benefit from cost-based optimizations, run periodic ANALYZE
statements on every large table that is frequently queried.
Fine-tuning#
The files_modified_after property is useful if you want to run the
ANALYZE statement on a table that was previously analyzed. You can use it to
limit the amount of data used to generate the table statistics:
ANALYZE example_table WITH(files_modified_after = TIMESTAMP '2021-08-23
16:43:01.321 Z')
As a result, only files newer than the specified time stamp are used in the analysis.
You can also specify a set or subset of columns to analyze using the columns
property:
ANALYZE example_table WITH(columns = ARRAY['nationkey', 'regionkey'])
To run ANALYZE with columns more than once, the next ANALYZE must
run on the same set or a subset of the original columns used.
To broaden the set of columns, drop the statistics and reanalyze the table.
Disable and drop extended statistics#
You can disable extended statistics with the catalog configuration property
delta.extended-statistics.enabled set to false. Alternatively, you can
disable it for a session, with the catalog session property extended_statistics_enabled set to false.
If a table is changed with many delete and update operation, calling ANALYZE
does not result in accurate statistics. To correct the statistics, you have to
drop the extended statistics and analyze the table again.
Use the system.drop_extended_stats procedure in the catalog to drop the
extended statistics for a specified table in a specified schema:
CALL example.system.drop_extended_stats('example_schema', 'example_table')
Memory usage#
The Delta Lake connector is memory intensive and the amount of required memory grows with the size of Delta Lake transaction logs of any accessed tables. It is important to take that into account when provisioning the coordinator.
You must decrease memory usage by keeping the number of active data files in
the table low by regularly running OPTIMIZE and VACUUM in Delta Lake.
Memory monitoring#
When using the Delta Lake connector, you must monitor memory usage on the coordinator. Specifically, monitor JVM heap utilization using standard tools as part of routine operation of the cluster.
A good proxy for memory usage is the cache utilization of Delta Lake caches. It
is exposed by the connector with the
plugin.deltalake.transactionlog:name=<catalog-name>,type=transactionlogaccess
JMX bean.
You can access it with any standard monitoring software with JMX support, or use the JMX connector with the following query:
SELECT * FROM jmx.current."*.plugin.deltalake.transactionlog:name=<catalog-name>,type=transactionlogaccess"
Following is an example result:
datafilemetadatacachestats.hitrate | 0.97
datafilemetadatacachestats.missrate | 0.03
datafilemetadatacachestats.requestcount | 3232
metadatacachestats.hitrate | 0.98
metadatacachestats.missrate | 0.02
metadatacachestats.requestcount | 6783
node | trino-master
object_name | io.trino.plugin.deltalake.transactionlog:type=TransactionLogAccess,name=delta
In a healthy system, both datafilemetadatacachestats.hitrate and
metadatacachestats.hitrate are close to 1.0.
Table redirection#
Trino offers the possibility to transparently redirect operations on an existing table to the appropriate catalog based on the format of the table and catalog configuration.
In the context of connectors which depend on a metastore service (for example, Hive connector, Iceberg connector and Delta Lake connector), the metastore (Hive metastore service, AWS Glue Data Catalog) can be used to accustom tables with different table formats. Therefore, a metastore database can hold a variety of tables with different table formats.
As a concrete example, let’s use the following simple scenario which makes use of table redirection:
USE example.example_schema;
EXPLAIN SELECT * FROM example_table;
Query Plan
-------------------------------------------------------------------------
Fragment 0 [SOURCE]
...
Output[columnNames = [...]]
│ ...
└─ TableScan[table = another_catalog:example_schema:example_table]
...
The output of the EXPLAIN statement points out the actual
catalog which is handling the SELECT query over the table example_table.
The table redirection functionality works also when using fully qualified names for the tables:
EXPLAIN SELECT * FROM example.example_schema.example_table;
Query Plan
-------------------------------------------------------------------------
Fragment 0 [SOURCE]
...
Output[columnNames = [...]]
│ ...
└─ TableScan[table = another_catalog:example_schema:example_table]
...
Trino offers table redirection support for the following operations:
Table read operations
Table write operations
Table management operations
Trino does not offer view redirection support.
The connector supports redirection from Delta Lake tables to Hive tables
with the delta.hive-catalog-name catalog configuration property.
Performance tuning configuration properties#
The following table describes performance tuning catalog properties specific to the Delta Lake connector.
Warning
Performance tuning configuration properties are considered expert-level features. Altering these properties from their default values is likely to cause instability and performance degradation. It is strongly suggested that you use them only to address non-trivial performance issues, and that you keep a backup of the original values if you change them.
Property name |
Description |
Default |
|---|---|---|
|
Minimum size of query predicates above which Trino compacts the predicates. Pushing a large list of predicates down to the data source can compromise performance. For optimization in that situation, Trino can compact the large predicates. If necessary, adjust the threshold to ensure a balance between performance and predicate pushdown. |
|
|
The target number of buffered splits for each table scan in a query, before the scheduler tries to pause. |
|
|
Sets the maximum number of splits used per second to access underlying storage. Reduce this number if your limit is routinely exceeded, based on your filesystem limits. This is set to the absolute maximum value, which results in Trino maximizing the parallelization of data access by default. Attempting to set it higher results in Trino not being able to start. |
|
|
For each query, the coordinator assigns file sections to read first
at the |
|
|
Sets the initial data size for a single read section
assigned to a worker until |
|
|
Sets the largest data size for a single read section
assigned to a worker after |
|
|
A decimal value in the range (0, 1] used as a minimum for weights assigned to each split. A low value might improve performance on tables with small files. A higher value might improve performance for queries with highly skewed aggregations or joins. |
|
|
Read only projected fields from row columns while performing |
|
|
Set to |
|