Starburst Snowflake connector#
The Snowflake connector allows querying and creating tables in an external Snowflake database. This can be used to join data between different systems like Snowflake and Hive, or between different Snowflake instances.
Requirements#
To connect to Snowflake, you need:
Network access from the coordinator and workers to Snowflake.
A valid Starburst Enterprise license.
In addition, the following line must be added to the jvm.config
file:
--add-opens=java.base/java.nio=ALL-UNNAMED
Configuration#
To configure the Snowflake connector, create a catalog properties file in
etc/catalog
named example.properties
, to mount the Snowflake connector
as the example
catalog (replace example
with your database name or some
other descriptive name of the catalog).
There are three connectors for Snowflake:
snowflake_jdbc
snowflake_parallel
snowflake_distributed
(deprecated)
snowflake_jdbc
uses JDBC for all reads and writes and is more efficient
when the result set returned from Snowflake is small.
We recommend using the snowflake_parallel
connector, which uses the JDBC
client to execute queries, and requests data in Arrow format. The connector then
reads each file independently from different nodes, which results in more
efficient querying.
When larger result sets are extracted from Snowflake, the
snowflake_distributed
connector may be a better choice. Instead of
requesting query results over a JDBC connection, the connector asks Snowflake to
export them to object storage and SEP reads them from there. Since write and
read operations are parallelized, this approach scales better for large data
sets, but has a higher latency.
Warning
The snowflake_distributed
connector is deprecated as of SEP 429-e. Use
the snowflake_parallel
connector instead, and migrate existing Snowflake catalogs to the parallel connector.
A requirement of using the distributed connector is that your Snowflake deployment runs on AWS or Azure. The connector automatically creates temporary stages in temporary storage buckets using Snowflake.
Create the catalog properties file, for example
etc/catalog/example.properties
, with the following contents, replacing the
connection properties as appropriate for your setup (for example, replace
<account_name>
with the full name of your account, as provided by
Snowflake).
For AWS:
connector.name=<snowflake_jdbc, snowflake_parallel, or snowflake_distributed>
connection-url=jdbc:snowflake://<account_name>.snowflakecomputing.com/
connection-user=<user_name>
connection-password=<password>
snowflake.warehouse=<warehouse_name>
snowflake.database=<database_name>
For Azure:
connector.name=<snowflake_jdbc, snowflake_parallel, or snowflake_distributed>
connection-url=jdbc:snowflake://<account_name>.<azure_region>.azure.snowflakecomputing.com/
connection-user=<user_name>
connection-password=<password>
snowflake.warehouse=<warehouse_name>
snowflake.database=<database_name>
The role used by Snowflake to execute operations can be specified as
snowflake.role=<role_name>
. This configuration is optional, and can not be
used together with User impersonation.
The catalog configuration specifies the warehouse used to execute queries with
the snowflake.warehouse
property. Use the catalog session property
warehouse
, if you want to temporarily switch to a different warehouse in the
current session for the user:
SET SESSION datacloud.warehouse = 'OTHER_WH';
If you switch to a non-existent warehouse, any following queries fail with an error message about missing an active warehouse.
You can verify the change to a warehouse with the following query:
SHOW SESSION LIKE '%warehouse';
This query either returns the session property with the name of the current warehouse in the session as value or no results, which signals that the warehouse from the catalog configuration is in use.
The Snowflake connector has the following configuration property:
Property name |
Description |
Default |
---|---|---|
|
Allow access to other databases in Snowflake by including the database name in double quotes with the schema name: SELECT *
FROM catalog."database.schema".table
When enabled, |
false |
Additionally, there are a number of configuration properties that apply only to the distributed connector:
Property name |
Description |
Default |
---|---|---|
|
Name of the schema in which stages are created for exporting data |
|
|
Number of export retries |
3 |
|
Maximum block size when reading from the export file |
16MB |
|
Maximum split size for processing the export file |
64MB |
|
Maximum initial split size |
Half of |
|
Maximum size of files to create when exporting data |
16MB |
Migrate to snowflake_parallel
#
To configure an existing Snowflake catalog to use the snowflake_parallel
connector, change the value of connector.name
in the catalog configuration
properties to snowflake_parallel
. The following catalog configuration
properties are not used by the parallel connector and must be removed:
snowflake.stage-schema
snowflake.max-export-retries
snowflake.parquet.max-read-block-size
snowflake.max-split-size
snowflake.max-initial-split-size
snowflake.export-file-max-size
The rest of the configuration is identical
to the snowflake_jdbc
connector.
General configuration properties#
The following table describes general catalog configuration properties for the connector:
Property name |
Description |
Default value |
---|---|---|
|
Support case insensitive schema and table names. |
|
|
This value should be a duration. |
|
|
Path to a name mapping configuration file in JSON format that allows Trino to disambiguate between schemas and tables with similar names in different cases. |
|
|
Frequency with which Trino checks the name matching configuration file for changes. This value should be a duration. |
(refresh disabled) |
|
The duration for which metadata, including table and column statistics, is cached. |
|
|
Cache the fact that metadata, including table and column statistics, is not available |
|
|
Maximum number of objects stored in the metadata cache |
|
|
Maximum number of statements in a batched execution. Do not change this setting from the default. Non-default values may negatively impact performance. |
|
|
Push down dynamic filters into JDBC queries |
|
|
Maximum duration for which Trino will wait for dynamic filters to be collected from the build side of joins before starting a JDBC query. Using a large timeout can potentially result in more detailed dynamic filters. However, it can also increase latency for some queries. |
|
Type mapping#
Because Trino and Snowflake each support types that the other does not, this connector modifies some types when reading or writing data. Data types may 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.
Snowflake to Trino type mapping#
The connector maps Snowflake types to the corresponding Trino types according to the following table:
Snowflake type |
Trino type |
Notes |
---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Supports only precision 3 in distributed mode |
|
|
Supports only precision 3 in distributed mode |
|
|
Supports only precision 3 in distributed mode |
|
|
Supports only precision 3 in distributed mode |
No other types are supported.
Trino to Snowflake type mapping#
The connector maps Trino types to the corresponding Snowflake types according to the following table:
Trino type |
Snowflake type |
Notes |
---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
No other types are supported.
Type mapping configuration properties#
The following properties can be used to configure how data types from the connected data source are mapped to Trino data types and how the metadata is cached in Trino.
Property name |
Description |
Default value |
---|---|---|
|
Configure how unsupported column data types are handled:
The respective catalog session property is |
|
|
Allow forced mapping of comma separated lists of data types to convert to
unbounded |
SQL support#
The connector provides read and write access to data and metadata in the Snowflake database. In addition to the globally available and read operation statements, the connector supports the following features:
SQL DELETE#
If a WHERE
clause is specified, the DELETE
operation only works if the
predicate in the clause can be fully pushed down to the data source.
ALTER TABLE EXECUTE#
The connector supports the following commands for use with ALTER TABLE EXECUTE:
collect_statistics#
The collect_statistics
command is used with
Managed statistics to collect statistics for a table
and its columns.
The following statement collects statistics for the example_table
table
and all of its columns:
ALTER TABLE example_table EXECUTE collect_statistics;
Collecting statistics for all columns in a table may be unnecessarily
performance-intensive, especially for wide tables. To only collect statistics
for a subset of columns, you can include the columns
parameter with an
array of column names. For example:
ALTER TABLE example_table
EXECUTE collect_statistics(columns => ARRAY['customer','line_item']);
Table functions#
The snowflake_jdbc
connector provides specific table functions to access Snowflake.
Note
The snowflake_distributed
connector does not support table functions.
Only snowflake_jdbc
and snowflake_parallel
connectors support table
functions.
query(VARCHAR) -> table
#
The query
function allows you to query the underlying database directly. It
requires syntax native to the data source, because the full query is pushed down
and processed in the data source. This can be useful for accessing native
features or for improving query performance in situations where running a query
natively may be faster.
The query
table function is available in the system
schema of any
catalog that uses the Snowflake connector, such as example
. The following
example passes myQuery
to the data source. myQuery
has to be a valid
query for the data source, and is required to return a table as a result:
SELECT
*
FROM
TABLE(
example.system.query(
query => 'myQuery'
)
);
Performance#
The connector includes a number of performance improvements, detailed in the following sections.
Table statistics#
The Snowflake connector reads table statistics, collected automatically by Snowflake, for cost based optimizations to improve query
processing performance based on the actual data in the data source. These table
statistics are read from Snowflake’s INFORMATION_SCHEMA.TABLES
table.
Note
The Snowflake connector can only collect table statistics. Column statistics
are not made available and therefore return as NULL
in a
SHOW STATS query. You can instead use managed statistics to have
SEP collect statistics from Snowflake, including column statistics.
Managed statistics#
The snowflake_jdbc
and snowflake_parallel
connectors support
Managed statistics allowing SEP to collect and store its own
table and column statistics that can then be used for performance optimizations
in query planning.
Statistics must be collected manually using the built-in
collect_statistics
command, see
collect_statistics for details and examples.
Pushdown#
The connector supports pushdown for a number of operations:
Aggregate pushdown for the following functions:
count()
, alsocount(distinct x)
avg(FLOAT)()
avg(DECIMAL)()
variance()
andvar_samp()
Join pushdown#
The join-pushdown.enabled
catalog configuration property or
join_pushdown_enabled
catalog session property control whether the connector pushes
down join operations. The property defaults to false
, and enabling join
pushdowns may negatively impact performance for some queries.
Join pushdown is set to false
by default in Snowflake catalogs to provide an
effective measure against higher compute costs.
JSON predicate pushdown#
Pushdown of json_extract()
and json_extract_scalar()
is considered
experimental, and must be enabled by setting the
snowflake.experimental-pushdown.enabled
catalog configuration property or
experimental_pushdown_enabled
catalog session property to true
.
Dynamic filtering#
Dynamic filtering is enabled by default. It causes the connector to wait for dynamic filtering to complete before starting a JDBC query.
You can disable dynamic filtering by setting the dynamic-filtering.enabled
property in your catalog configuration file to false
.
Wait timeout#
By default, table scans on the connector are delayed up to 20 seconds until dynamic filters are collected from the build side of joins. Using a large timeout can potentially result in more detailed dynamic filters. However, it can also increase latency for some queries.
You can configure the dynamic-filtering.wait-timeout
property in your
catalog properties file:
dynamic-filtering.wait-timeout=1m
You can use the dynamic_filtering_wait_timeout
catalog session property in a specific session:
SET SESSION catalogname.dynamic_filtering_wait_timeout = 1s;
Compaction#
The maximum size of dynamic filter predicate, that is pushed down to the
connector during table scan for a column, is configured using the
domain-compaction-threshold
property in the catalog
properties file:
domain-compaction-threshold=100
You can use the domain_compaction_threshold
catalog
session property:
SET SESSION domain_compaction_threshold = 10;
By default, domain-compaction-threshold
is set to 32
.
When the dynamic predicate for a column exceeds this threshold, it is compacted
into a single range predicate.
For example, if the dynamic filter collected for a date column dt
on the
fact table selects more than 32 days, the filtering condition is simplified from
dt IN ('2020-01-10', '2020-01-12',..., '2020-05-30')
to dt BETWEEN '2020-01-10' AND '2020-05-30'
. Using a large threshold can result in increased
table scan overhead due to a large IN
list getting pushed down to the data
source.
Metrics#
Metrics about dynamic filtering are reported in a JMX table for each catalog:
jmx.current."com.starburstdata.presto.plugin.jdbc.dynamicfiltering:catalog=snowflake,name=snowflake,type=dynamicfilteringstats"
Metrics include information about the total number of dynamic filters, the number of completed dynamic filters, the number of available dynamic filters and the time spent waiting for dynamic filters.
JDBC connection pooling#
When JDBC connection pooling is enabled, each node creates and maintains a connection pool instead of opening and closing separate connections to the data source. Each connection is available to connect to the data source and retrieve data. After completion of an operation, the connection is returned to the pool and can be reused. This improves performance by a small amount, reduces the load on any required authentication system used for establishing the connection, and helps avoid running into connection limits on data sources.
JDBC connection pooling is disabled by default. You can enable JDBC connection
pooling by setting the connection-pool.enabled
property to true
in your
catalog configuration file:
connection-pool.enabled=true
The following catalog configuration properties can be used to tune connection pooling:
Property name |
Description |
Default value |
---|---|---|
|
Enable connection pooling for the catalog. |
|
|
The maximum number of idle and active connections in the pool. |
|
|
The maximum lifetime of a connection. When a connection reaches this lifetime it is removed, regardless of how recently it has been active. |
|
|
The maximum size of the JDBC data source cache. |
|
|
The expiration time of a cached data source when it is no longer accessed. |
|
Starburst Cached Views#
The connectors supports table scan redirection to improve performance and reduce load on the data source.
Security#
The connector includes a number of security-related features, detailed in the following sections.
User impersonation#
The Snowflake connector supports User impersonation. It can be
configured to use a number of different impersonation mechanisms, specified by
the values configured for the property snowflake.impersonation-type
:
snowflake.impersonation-type=ROLE
NONE
(default value)Connect as the service user with the credentials from the catalog properties file and assume the role defined by the
snowflake.role
property, or the service user’s default role, if the property is missing.ROLE
Connect as the service user and use auth-to-local mapping to map the user to a role.
OKTA_LDAP_PASSTHROUGH
Assume the identity of the SEP user, authenticated with Okta using LDAP credentials, and use the default Snowflake user role.
ROLE_OKTA_LDAP_PASSTHROUGH
As above, but additionally use auth-to-local mapping to map the user to a role.
Note
Impersonation using Okta is not supported for Azure.
Authentication with Okta#
The Snowflake connector supports the usage of the Okta Single Sign-On system to authenticate users via SEP to Snowflake.
Note
Okta authentication is not supported for Azure.
The setup allows users to authenticate to SEP using their LDAP credentials and use the same credentials to authenticate to Snowflake through Okta. The same credentials are used by SEP when accessing data in Snowflake.
Behind the scenes SEP and the Snowflake connector authenticate to Okta with the LDAP credentials of the user. After the user authenticates with Okta, including MFA potentially, a SAML assertion allows Snowflake to issue an an OAuth 2.0 token pair. The tokens are cached in SEP and used for further authentications until they expire, and another authentication is requested.
If Okta multi-factor authentication (MFA) is configured, users have to confirm authentication with it. One time codes are not supported.
To enable the Okta integration, SEP and Snowflake need to be configured correctly.
Okta and SEP are configured to use LDAP authentication using the same user
identifiers and LDAP directory. In addition to the usual LDAP configuration,
LDAP authentication, for SEP, you need to enable
Password credential pass-through in etc/config.properties
:
http-server.authentication.type=DELEGATED-PASSWORD
Snowflake is configured in Okta as a SAML application as detailed in the
Snowflake
documentation.
Note that the Snowflake login_name
must match the corresponding SAML Subject
NameID
attribute value.
SEP is configured as an OAuth client in Snowflake, again detailed in the Snowflake OAuth documentation.
Note
You must pre-authorize user consent for a role to allow authentication to occur without user interaction.
A number of properties are required to configure the Okta integration in the Snowflake catalog properties file:
snowflake.account-name
Name of Snowflake account.
snowflake.account-url
URL of the Snowflake account. The URL usually has the form https://account_name.snowflakecomputing.com, but might include additional segments.
snowflake.client-id
Snowflake OAuth client id. This can be retrieved with the secret name and a query like
select system$show_oauth_client_secrets('OAUTH_TEST_INT');
.snowflake.client-secret
Snowflake OAuth client secret.
snowflake.credential.cache-ttl
Duration the OAuth refresh token is cached. This value cannot exceed the
oauth_refresh_token_validity
value used when the OAuth integration was created. E.g.24h
.okta.account-url
The Okta URL, typically https://your_okta_account_name.okta.com).
Optional properties allow you to override the default values:
snowflake.credential.cache-size
The size of the OAuth credentials cache. Use a value that accommodates the expected number of users that might connect to Snowflake through SEP during the period defined by the TTL of the token. Defaults to 10000.
snowflake.credential.http-connect-timeout
Connection timeout. Defaults to 30s.
snowflake.credential.http-read-timeout
Connection read timeout. Defaults to 30s.
snowflake.credential.http-write-timeout
Connection write timeout. Defaults to 30s.
snowflake.redirect-uri
The redirect URI for OAUTH. Value must match the redirect URI specified when creating the security integration (oauth_redirect_uri). Defaults to
https://localhost
.okta.credential.http-connect-timeout
Connection timeout. Defaults to 30s.
okta.credential.http-read-timeout
Connection read timeout. Defaults to 30s.
okta.credential.http-write-timeout
Connection write timeout. Defaults to 30s.
OAuth 2.0 token pass-through#
The Snowflake connector supports OAuth 2.0 token pass-through.
Configure this option the same as Authentication with Okta, except for the settings described in this section.
Set the authentication type and OAuth 2.0 scope in the coordinator’s config properties file:
http-server.authentication.type=DELEGATED-OAUTH2
http-server.authentication.oauth2.scopes=<EXISTING_SCOPES>,session:role:TEST_ROLE
The session:role
prefix determines the role assigned to the user after
successful authentication.
Additionally enable OAUTH2_PASSTHROUGH
in the catalog properties file using
the Snowflake connector:
snowflake.impersonation-type=OAUTH2_PASSTHROUGH