Database engine specifications

Superset uses SQLAlchemy as an abstraction layer for running queries and fetching metadata from tables (like column names and types). Unfortunately, while SQLAlchemy offers enough functionality to allow connecting Superset to dozens of databases, there are still implementation details that differ across them. Because of this, Superset has an additional abstraction on top of SQLAlchemy, called a “database engine specification” or, simply, “DB engine spec”.

DB engine specs were created initially because there's no SQL standard for computing aggregations at different time grains. For example, to compute a daily metric in Trino or Postgres we could run a query like this:

SELECT
  date_trunc('day', CAST(time_column) AS TIMESTAMP) AS day,
  COUNT(*) AS metric
FROM
  some_table
GROUP BY
  1

For MySQL, instead of using the date_trunc function, we would need to write:

SELECT
  DATE(time_column) AS day,
  COUNT(*) AS metric
FROM
  some_table
GROUP BY
  1

Over time, more and more functionality was added to DB engine specs, including validating SQL, estimating the cost of queries before they are run, and understanding the semantics of error messages. These are all described in detail in this document, and in the table below you can see a summary of what features are supported by each database.

Note that DB engine specs are completely optional. Superset can connect to any database supported by SQLAlchemy (or 3rd party dialects) even if there's no DB engine spec associated with it. But DB engine specs greatly improve the experience of working with a database in Superset.

Features

The table below (generated via python superset/db_engine_specs/lib.py) summarizes the information about the status of all DB engine specs in Superset (note that this excludes 3rd party DB engine specs):

(Note, this table is generated via: python superset/db_engine_specs/lib.py.)

Database information

A DB engine spec has attributes that describe the underlying database engine, so that Superset can know how to build and run queries. For example, some databases don‘t support subqueries, which are needed for some of the queries produced by Superset for certain charts. When a database doesn’t support subqueries the query is run in two-steps, using the results from the first query to build the second query.

These attributes and their default values (set in the base class, BaseEngineSpec) are described below:

limit_method = LimitMethod.FORCE_LIMIT

When running user queries in SQL Lab, Superset needs to limit the number of rows returned. The reason for that is cost and performance: there‘s no point in running a query that produces millions of rows when they can’t be loaded into the browser.

For most databases this is done by parsing the user submitted query and applying a limit, if one is not present, or replacing the existing limit if it's larger. This is called the FORCE_LIMIT method, and is the most efficient, since the database will produce at most the number of rows that Superset will display.

For some databases this method might not work, and they can use the WRAP_SQL method, which wraps the original query in a SELECT * and applies a limit via the SQLAlchemy dialect, which should get translated to the correct syntax. This method might be inefficient, since the database optimizer might not be able to push the limit to the inner query.

Finally, as a last resource there is the FETCH_MANY method. When a DB engine spec uses this method the query runs unmodified, but Superset fetches only a certain number of rows from the cursor. It‘s possible that a database using this method can optimize the query execution and compute rows as they are being read by the cursor, but it’s unlikely. This makes this method the least efficient of the three.

Note that when Superset runs a query with a given limit, say 100, it always modifies the query to request one additional row (LIMIT 101, in this case). This extra row is dropped before the results are returned to the user, but it allows Superset to inform the users that the query was indeed limited. Otherwise a query with LIMIT 100 that returns exactly 100 rows would seem like it was limited, when in fact it was not.

allows_joins = True

Not all databases support JOINs. When building complex charts, Superset will try to join the table to itself in order to compute top_n groups, for example. If the database doesn't support joins Superset will instead run a prequery, and use the results to build the final query.

allows_subqueries = True

Similarly, not all databases support subqueries. For more complex charts Superset will build subqueries if possible, or run the query in two-steps otherwise.

allows_alias_in_select = True

Does the DB support aliases in the projection of a query, eg:

SELECT COUNT(*) AS cnt

Superset will try to use aliases whenever possible, in order to give friendly names to expressions.

allows_alias_in_orderby = True

Does the DB support referencing alias in the GROUP BY, eg:

SELECT
  UPPER(country_of_origin) AS country
  COUNT(*) AS cnt
FROM
  some_table
GROUP BY
  country

Otherwise the query is written as:

SELECT
  UPPER(country_of_origin) AS country
  COUNT(*) AS cnt
FROM
  some_table
GROUP BY
  UPPER(country_of_origin)

time_groupby_inline = False

In theory this attribute should be used to omit time filters from the self-joins. When the attribute is false the time attribute will be present in the subquery used to compute limited series, eg:

SELECT DATE_TRUNC('day', ts) AS ts,
       team AS team,
       COUNT(*) AS count
FROM public.threads
JOIN
  (SELECT team AS team__,
          COUNT(*) AS mme_inner__
   FROM public.threads
   -- this is added when `time_groupby_inline = False`
   WHERE ts >= TO_TIMESTAMP('2022-07-27 00:00:00.000000', 'YYYY-MM-DD HH24:MI:SS.US')
     AND ts < TO_TIMESTAMP('2023-07-27 00:00:00.000000', 'YYYY-MM-DD HH24:MI:SS.US')
   --
   GROUP BY team
   ORDER BY mme_inner__ DESC
   LIMIT 5) AS anon_1 ON team = team__
WHERE ts >= TO_TIMESTAMP('2022-07-27 00:00:00.000000', 'YYYY-MM-DD HH24:MI:SS.US')
  AND ts < TO_TIMESTAMP('2023-07-27 00:00:00.000000', 'YYYY-MM-DD HH24:MI:SS.US')
GROUP BY DATE_TRUNC('day', ts),
         team
ORDER BY count DESC
LIMIT 10000;

In practice, the attribute doesn't seem to be working as of 2023-07-27.

allows_alias_to_source_column = True

When this is true the database allows queries where alias can overshadow existing column names. For example, in this query:

SELECT
  foo + 1 AS foo
FROM
  some_table
ORDER BY
  foo  -- references the alias `foo + 1`, not the column `foo`

allows_hidden_orderby_agg = True

If set to true the database allows expressions in the GROUP BY that are not present in the projection (SELECT), eg:

SELECT
  country,
  COUNT(*)
FROM
  some_table
GROUP BY
  country
ORDER BY
  SUM(population)  -- not present in the `SELECT`

allows_hidden_cc_in_orderby = False

This the opposite of allows_alias_in_orderby, for databases that require aliases in the ORDER BY. For example, BigQuery doesn't like this query:

SELECT
  CASE
    WHEN type = 'feature' THEN 'f'
    WHEN type = 'bug' THEN 'b'
    ELSE 'o'
  END AS cc_type
FROM
  some_table
GROUP BY
  cc_type
ORDER BY
  CASE
    WHEN type = 'feature' THEN 'f'
    WHEN type = 'bug' THEN 'b'
    ELSE 'o'
  END

Instead, it must be written as:

SELECT
  CASE
    WHEN type = 'feature' THEN 'f'
    WHEN type = 'bug' THEN 'b'
    ELSE 'o'
  END AS cc_type
FROM
  some_table
GROUP BY
  cc_type
ORDER BY
  cc_type

allows_cte_in_subquery = True

When a virtual dataset is used in a chart the original query is converted into a subquery, and is wrapped in an outer query that is generated based on the chart controls. The virtual dataset query might have a CTE, and some databases don't like subqueries with CTEs in them.

When this attribute is false Superset will extract the CTE and move it outside of the subquery when generating SQL for charts. The name of the new CTE will be cte_alias, also defined in the DB engine spec.

allow_limit_clause = True

Allows for the LIMIT clause. Otherwise, the database probably uses TOP to limit rows.

max_column_name_length: int | None = None

Most databases have a well defined limit for the maximum length of a column name (SQLite is probably the one exception). While the can be set (and defaults) to None, it's highly recommended to set a value to prevent errors.

allows_sql_comments = True

Are comments supported in the DB? In general SQL in comments are defined by double dashes:

-- this is a comment
SELECT *  -- we need everything
FROM some_table

allows_escaped_colons = True

SQLAlchemy recommends escaping colons to prevent them from being interpreted as bindings to parameters. Because of this, when building queries from virtual datasets Superset will escape all colons with \:.

This works for most databases except Athena. The allows_escaped_colons attribute specifies if the database supports the escape colon.

Basic features

These are features that all DB engine specs should support, as the name suggests. They provide a much better user experience for the user.

Time grains

The most basic feature that DB engine specs need to support is defining time grain expressions. These are dialect-specific SQL expressions that are used to compute metrics on a given time grain when building charts. For example, when computing the metric COUNT(*) on a daily basis, Superset will generate the following query:

SELECT
  <DB engine spec expression for TimeGrain.DAY>,
  COUNT(*)
...
GROUP BY
  <DB engine spec expression for TimeGrain.DAY>

For some databases with support for DATE_TRUNC or TIME_FLOOR this is easy. Here's how Apache Druid computes 15 minute aggregations:

TIME_FLOOR(CAST({col} AS TIMESTAMP), 'PT15M')

Where {col} is the time column being aggregated — the expression is actually a Jinja2 template. Druid uses the ISO standard for durations, with PT15M representing 15 minutes.

On the other and, here's the same for SQLite:

DATETIME(
  STRFTIME(
    '%Y-%m-%dT%H:%M:00',
    {col}
  ),
  printf(
    '-%d minutes',
    CAST(strftime('%M', {col}) AS INT) % 15
  )
)

The SQLite version has to truncate the column down to the minute, and then subtract a number of minutes equals to the modulo 15.

Time grain expressions are defined in the _time_grain_expressions class attribute, which maps from a superset.constants.TimeGrain to the SQL expression. The dictionary has a special key None, that should map to the column directly, for when no time grain is specified.

Note that it's possible to add new time grains via configuration. For example, if you want to add a “2 seconds” time grain to your installation you can add it to TIME_GRAIN_ADDONS, and implement it in TIME_GRAIN_ADDON_EXPRESSIONS:

# superset_config.py
TIME_GRAIN_ADDONS = {"PT2S": "2 second"}

TIME_GRAIN_ADDON_EXPRESSIONS = {
    "clickhouse": {
        "PT2S": "toDateTime(intDiv(toUInt32(toDateTime({col})), 2)*2)",
    }
}

Column type mapping

Column type mapping, defined in the column_type_mappings class attribute, is just a way of mapping type names from the database to types Superset understand. The default values in BaseEngineSpec are sane:

_default_column_type_mappings: tuple[ColumnTypeMapping, ...] = (
    (
        re.compile(r"^string", re.IGNORECASE),
        types.String(),
        GenericDataType.STRING,
    ),
    (
        re.compile(r"^float", re.IGNORECASE),
        types.Float(),
        GenericDataType.NUMERIC,
    ),
    (
        re.compile(r"^date", re.IGNORECASE),
        types.Date(),
        GenericDataType.TEMPORAL,
    ),
    (
        re.compile(r"^bool(ean)?", re.IGNORECASE),
        types.Boolean(),
        GenericDataType.BOOLEAN,
    ),
    ...
)

But you might want to implement more specific types in the DB engine spec, or complex types. For example, for MSSQL we have:

from sqlalchemy.dialects.mssql.base import SMALLDATETIME

class MssqlEngineSpec(BaseEngineSpec):
    ...
    column_type_mappings = (
        (
            re.compile(r"^smalldatetime.*", re.IGNORECASE),
            SMALLDATETIME(),
            GenericDataType.TEMPORAL,
        ),
    )

Function names

DB engine specs should implement a class method called get_function_names that returns a list of strings, representing all the function names that the database supports. This is used for autocomplete in SQL Lab.

Masked encrypted extra

Superset does a good job in keeping credentials secure. When you add a database with a password, for example:

postgresql://admin:password123@db.example.org:5432/db

The password is sent over the network only when the database is created. When you edit the database later, Superset will return this as the SQLAlchemy URI:

postgresql://admin:XXXXXXXXXX@db.example.org:5432/db

The password will be masked in the API response; it‘s not just masked in the browser UI. This is done in order to avoid sending the password unnecessarily over the network. Also, if a non-admin user has access to the API response, they won’t be able to know the database password.

When the database is edited, the Superset backend is smart enough to replace the masked password with the actual password, unless the password has changed. That is, if you change the database in the URI from db to db2 the SQLAlchemy URI will be stored in the backend as:

postgresql://admin:password123@db.example.org:5432/db2

The password is not the only piece of information where security is critical. For many databases (like BigQuery), sensitive information is stored in the credentials JSON payload. For example:

{
  "type": "service_account",
  "project_id": "dbt-tutorial-347100",
  "private_key_id": "4bc71f06990c864a590fad8b94be6a5904fc171f",
  "private_key": "<SENSITIVE INFORMATION>",
  "client_email": "dbt-user-278@dbt-tutorial-347100.iam.gserviceaccount.com",
  "client_id": "115666988796889519425",
  "auth_uri": "https://accounts.google.com/o/oauth2/auth",
  "token_uri": "https://oauth2.googleapis.com/token",
  "auth_provider_x509_cert_url": "https://www.googleapis.com/oauth2/v1/certs",
  "client_x509_cert_url": "https://www.googleapis.com/robot/v1/metadata/x509/dbt-user-278%40dbt-tutorial-347100.iam.gserviceaccount.com"
}

Similarly to password, we don't want to send private_key to the client when a database is edited; the Superset API should never return its actual contents. Instead, Superset should return a masked value, and users should be able to edit the JSON without having to type in the private_key on every edit.

To do this, DB engine specs and implement 2 methods, mask_encrypted_extra and unmask_encrypted_extra. They have these names because the credentials are stored in an encrypted column called encrypted_extra. Here's how these methods look like for BigQuery:

from superset.constants import PASSWORD_MASK


class BigQueryEngineSpec(BaseEngineSpec):

    @classmethod
    def mask_encrypted_extra(cls, encrypted_extra: str | None) -> str | None:
        if encrypted_extra is None:
            return encrypted_extra

        try:
            config = json.loads(encrypted_extra)
        except (json.JSONDecodeError, TypeError):
            return encrypted_extra

        try:
            config["credentials_info"]["private_key"] = PASSWORD_MASK
        except KeyError:
            pass

        return json.dumps(config)

    @classmethod
    def unmask_encrypted_extra(
        cls,
        old: str | None,
        new: str | None
    ) -> str | None:
        if old is None or new is None:
            return new

        try:
            old_config = json.loads(old)
            new_config = json.loads(new)
        except (TypeError, json.JSONDecodeError):
            return new

        if "credentials_info" not in new_config:
            return new

        if "private_key" not in new_config["credentials_info"]:
            return new

        if new_config["credentials_info"]["private_key"] == PASSWORD_MASK:
            new_config["credentials_info"]["private_key"] = old_config[
                "credentials_info"
            ]["private_key"]

        return json.dumps(new_config)

This way, when a user edits an existing BigQuery connection, the private_key is shown as XXXXXXXXXX. Everything else in the JSON is still displayed, and the user can change any of the fields without having to provide the private key.

Note that while this is a basic feature that should be implemented for security reasons, it only makes sense in DB engine specs that use encrypted_extra to store connection information.

Nice to have features

The next set of features are nice to have. They don't apply to all databases, and are not strictly needed for security or usability.

User impersonation

In general there's no user-level granularity when accessing a database in Superset. A single database connection is shared by all users who have access to that database. There are many use cases when this is not desirable, and some databases implement mechanisms in which they can impersonate users, potentially reducing the scope of permissions available to run the query.

For example, the Google Sheets DB engine spec implements this via the get_url_for_impersonation class method:

class GSheetsEngineSpec(ShillelaghEngineSpec):

    @classmethod
    def get_url_for_impersonation(
        cls,
        url: URL,
        impersonate_user: bool,
        username: str | None,
    ) -> URL:
        if impersonate_user and username is not None:
            user = security_manager.find_user(username=username)
            if user and user.email:
                url = url.update_query_dict({"subject": user.email})

        return url

The method get_url_for_impersonation updates the SQLAlchemy URI before every query. In this particular case, it will fetch the user's email and add it to the subject query argument. The driver will then lower the permissions to match that given user. This allows the connection to be configured with a service account that has access to all the spreadsheets, while giving users access to only the spreadsheets they own are have been shared with them (or with their organization — Google will handle the authorization in this case, not Superset).

Alternatively, it's also possible to impersonate users by implementing the update_impersonation_config. This is a class method which modifies connect_args in place. You can use either method, and ideally they should be consolidated in a single one.

File upload

When a DB engine spec supports file upload it declares so via the supports_file_upload class attribute. The base class implementation is very generic and should work for any database that has support for CREATE TABLE. It leverages Pandas and the df_to_sql method.

For some databases the df_to_sql classmethod needs to be implemented. For example, for BigQuery the DB engine spec implements a custom method that uses the to_gbq method.

Extra table metadata

DB engine specs can return additional metadata associated with a table. This is done via the extra_table_metadata class method. Trino uses this to return information about the latest partition, for example, and Bigquery returns clustering information. This information is then surfaced in the SQL Lab UI, when browsing tables in the metadata explorer (on the left panel).

DB API exception mapping

Different DB API 2.0 drivers implement different exceptions, even if they have the same name. The get_dbapi_exception_mapping class method returns a dictionary mapping these custom exceptions to Superset exceptions, so that Superset can return more specific errors when an exception is raised by the underlying driver.

For example, for ClickHouse we have:

from urllib3.exceptions import NewConnectionError

from superset.db_engine_specs.exceptions import SupersetDBAPIDatabaseError


class ClickHouseEngineSpec(ClickHouseBaseEngineSpec):

    @classmethod
    def get_dbapi_exception_mapping(cls) -> dict[type[Exception], type[Exception]]:
        return {NewConnectionError: SupersetDBAPIDatabaseError}

This way, if the ClickHouse driver raises a NewConnectionError it would get wrapped in a SupersetDBAPIDatabaseError.

Custom errors

Queries can fail in many different ways. For example, in SQLite:

sqlite> CREATE TABLE a (b INT);
sqlite> SELECT c FROM a;
Error: no such column: c
sqlite>

When a query fails, Superset will return the message, “Error: no such column: c”, to the user as a generic error.

Since ideally we want to return specific and actionable error messages, DB engine specs can implement methods that map error messages to more specific errors. For example, the SQLite DB engine specs defines:

COLUMN_DOES_NOT_EXIST_REGEX = re.compile("no such column: (?P<column_name>.+)")


class SqliteEngineSpec(BaseEngineSpec):

    custom_errors: dict[Pattern[str], tuple[str, SupersetErrorType, dict[str, Any]]] =
        COLUMN_DOES_NOT_EXIST_REGEX: (
            __('We can\'t seem to resolve the column "%(column_name)s"'),
            SupersetErrorType.COLUMN_DOES_NOT_EXIST_ERROR,
            {},
        ),
    }

This way, when a user selects a column that doesn't exist Superset can return a more informative error.

Dynamic schema

In SQL Lab it's possible to select a database, and then a schema in that database. Ideally, when running a query in SQL Lab, any unqualified table names (eg, table, instead of schema.table) should be in the selected schema. For example, if the user select dev as the schema and then runs the following query:

SELECT * FROM my_table

The table my_table should live in the dev schema. In order to do that, it's necessary to modify the SQLAlchemy URI before running the query. Since different databases have different ways of doing that, this functionality is implemented via the adjust_engine_params class method. The method receives the SQLAlchemy URI and connect_args, as well as the schema in which the query should run. It then returns a potentially modified URI and connect_args to ensure that the query runs in the specified schema.

When a DB engine specs implements adjust_engine_params it should have the class attribute supports_dynamic_schema set to true. This is critical for security, since it allows Superset to know to which schema any unqualified table names belong to. For example, in the query above, if the database supports dynamic schema, Superset would check to see if the user running the query has access to dev.my_table. On the other hand, if the database doesn't support dynamic schema, Superset would sue the default database schema instead of dev.

Implementing this method is also important for usability. When the method is not implemented selecting the schema in SQL Lab has no effect on the schema in which the query runs, resulting in a confusing results when using unqualified table names.

Catalog

In general, databases support a hierarchy of concepts of one-to-many concepts:

1. Database
2. Catalog
3. Namespace
4. Table
5. Column

These concepts have different names depending on the database. For example, Postgres uses the following terminology:

1. Cluster (database)
2. Database (catalog)
3. Schema (namespace)
4. Table
5. Column

BigQuery, on the other hand:

1. Bigquery (database)
2. Project (catalog)
3. Schema (namespace)
4. Table
5. Column

Hive and Trino:

1. Database
2. Catalog
3. Schema
4. Table
5. Column

If the database supports catalogs, then the DB engine spec should have the supports_catalog class attribute set to true.

Dynamic catalog

Superset has no support for multiple catalogs. A given SQLAlchemy URI connects to a single catalog, and it's impossible to browse other catalogs, or change the catalog. This means that datasets can only be added for the main catalog of the database. For example, with this Postgres SQLAlchemy URI:

postgresql://admin:password123@db.example.org:5432/db

Here, datasets can only be added to the db catalog (which Postgres calls a “database”).

One confusing problem is that many databases allow querying across catalogs in SQL Lab. For example, with BigQuery one can write:

SELECT * FROM project.schema.table

This means that even though the database is configured for a given catalog (project), users can query other projects. This is a common workaround for creating datasets in catalogs other than the catalog configured in the database: just create a virtual dataset.

Ideally we would want users to be able to choose the catalog when using SQL Lab and when creating datasets. In order to do that, DB engine specs need to implement a method that rewrites the SQLAlchemy URI depending on the desired catalog. This method already exists, and is the same method used for dynamic schemas, adjust_engine_params, but currently there are no UI affordances for choosing a catalog.

Before the UI is implemented Superset still needs to implement support for catalogs in its security manager. But in the meantime, it's possible for DB engine spec developers to support dynamic catalogs, by setting supports_dynamic_catalog to true and implementing adjust_engine_params to handle a catalog.

SSH tunneling

Superset can connect to databases via an SSH tunnel. For databases where this doesn't make sense (eg, SQLite or BigQuery) the DB engine spec should have disable_ssh_tunneling set to true.

Query cancelation

Superset will try to cancel running queries if the users wants so, but it's up to the DB engine spec to handle this.

Some databases have an implicit query cancelation. When a cursor stops being polled it will cancel the query. For databases that behave like this, the class method has_implicit_cancel (which should really be a class attribute) should return true.

For other databases, DB engine specs can implement query cancelation via the prepare_cancel_query and cancel_query methods. Implementation of query cancelation is usually heavily dependent on the database, but the DB engine specs that support it can serve as an example.

Get metrics on dataset creation

When a physical dataset is first created, the get_metrics class method is called on the table. The base implementation returns the COUNT(*) metric, but DB engine specs can override get_metrics to return other metrics. This method is useful for semantic layers that contain their own metrics definitions; when Superset connect to them it can automatically create those metrics when a dataset is added.

This feature is still experimental, and ideally there would be a mechanism for calling it periodically or when a dataset is explored, in order to sync new metric definitions to the dataset.

WHERE on latest partition

In some databases, running SELECT * can be a very expensive operation, since the query might scan all partitions for a given table. Because of that, some DB engine specs implement the where_latest_partition method, which returns a modified SQLAlchemy query with an additional predicate that filters on the latest partition.

Advanced features

Expand complex types

Some databases will visually expand complex types (arrays and structures) when displaying results from queries. For example, the BigQuery UI is able to expand objects into columns and array into rows, so that this:

Is shown as:

A similar behavior has been implemented in Superset for Presto, and can be enabled via the PRESTO_EXPAND_DATA feature flag. To implement this feature a DB engine spec should implement the expand_data method, which takes the columns and rows and returns modified columns and rows.

Note that despite being implemented only for Presto, this behavior has nothing that is Presto specific, and in theory could be implemented in a generic way for all database without requiring custom DB engine spec implementations (that is, the Presto expand_data method could be moved to the base class, after being cleaned up, and we could then enable the feature per DB in the configuration).

Query cost estimation

Some databases allow uses to estimate the cost of running a query before running it. This is done via the estimate_query_cost method in DB engine specs, which receives the SQL and returns a list of “costs”. The definition of what “cost” is varies from database to database (in the few that support this functionality), and it can be formatted via the query_cost_formatter.

The query_cost_formatter can be overridden with an arbitrary function via the config QUERY_COST_FORMATTERS_BY_ENGINE. This allows custom deployments of Superset to format the results in different ways. For example, at some point in Lyft the cost for running Presto queries would also show the carbon footprint (in trees).

SQL validation

A few databases support validating the syntax of the SQL as the user is typing it, indicating in SQL Lab any errors. This is usually done using an EXPLAIN query and, because it gets called every few seconds as the user types, it's important that the database returns the result quickly.

This is currently implement for Presto and Postgres, via custom classes in superset/sql_validators that should be enabled in the configuration. Implementing this as custom classes, instead of a validate_sql method in the DB engine spec offers no advantages, and ideally in the future we should move the logic to DB engine specs.

Testing DB engine specs

Superset has a command to test the connection to a given database, as well as checking if the SQLAlchemy dialect implements all necessary methods used by Superset, and checking which features are supported by the DB engine spec (if one exists). To run the tool just call the test-db command with the SQLAlchemy URI to be tested:

superset test-db sqlite://

If the connection needs additional arguments they can be passed when the command runs.