[
https://issues.apache.org/jira/browse/FLINK-5568?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15879942#comment-15879942
]
Kurt Young commented on FLINK-5568:
-----------------------------------
Hi [~fhueske],
I want to emphasize there exists a very big difference between this catalog
design and existing catalog like HCatalog or Hive. The difference is how we
answer the following question: "Does the table type which flink understands
bind with the catalog?". Why we need to answer this is because some catalog's
implementation has hidden the table details for us. Like HCatalog, it provides
a unified loader and storer layer to let you read and write all kinds of files,
like Parquet, ORC and so on. So, if there exists a table from HCatalog which
consists of parquet files, should we treat it as a "HCatalogTableSource", or
just "ParquetTableSource".
We think it's not a very good idea to have Flink's table bound with each
catalog, so we introduce another converting layer to transfer the table type.
This is actually a more flexible way to do since we can easily add
"HCatalogTableSource" to handle all the tables from HCatalog. And it's easier
for advanced users to implement their own catalog, the just need to tell Flink
what's type of the table, some necessary informations as well as a pluggable
"TableSource" implementation.
Let us know if you have some other thoughts.
> Introduce interface for catalog, and provide an in-memory implementation, and
> integrate with calcite schema
> -----------------------------------------------------------------------------------------------------------
>
> Key: FLINK-5568
> URL: https://issues.apache.org/jira/browse/FLINK-5568
> Project: Flink
> Issue Type: Sub-task
> Components: Table API & SQL
> Reporter: Kurt Young
> Assignee: jingzhang
>
> The {{TableEnvironment}} now provides a mechanism to register temporary
> table. It registers the temp table to calcite catalog, so SQL and TableAPI
> queries can access to those temp tables. Now DatasetTable, DataStreamTable
> and TableSourceTable can be registered to {{TableEnvironment}} as temporary
> tables.
> This issue wants to provides a mechanism to connect external catalogs such as
> HCatalog to the {{TableEnvironment}}, so SQL and TableAPI queries could
> access to tables in the external catalogs without register those tables to
> {{TableEnvironment}} beforehand.
> First, we should point out that there are two kinds of catalog in Flink
> actually.
> The first one is external catalog as we mentioned before, it provides CRUD
> operations to databases/tables.
> The second one is calcite catalog, it defines namespace that can be accessed
> in Calcite queries. It depends on Calcite Schema/Table abstraction.
> SqlValidator and SqlConverter depends on the calcite catalog to fetch the
> tables in SQL or TableAPI.
> So we need to do the following things:
> 1. introduce interface for external catalog, maybe provide an in-memory
> implementation first for test and develop environment.
> 2. introduce a mechanism to connect external catalog with Calcite catalog so
> the tables/databases in external catalog can be accessed in Calcite catalog.
> Including convert databases of externalCatalog to Calcite sub-schemas,
> convert tables in a database of externalCatalog to Calcite tables (only
> support {{TableSourceTable}}).
> 3. register external catalog to {{TableEnvironment}}.
> Here is the design mode of ExternalCatalogTable.
> | identifier | TableIdentifier | dbName and tableName
> of table |
> | tableType | String | type of external catalog table,
> e.g csv, hbase, kafka
> | schema | DataSchema| schema of table data,
> including column names and column types
> | partitionColumnNames | List<String> | names of partition column
> | properties | Map<String, String> |properties of
> external catalog table
> | stats | TableStats | statistics of external
> catalog table
> | comment | String |
> | create time | long
> There is still a detail problem need to be take into consideration, that is ,
> how to convert {{ExternalCatalogTable}} to {{TableSourceTable}}. The
> question is equals to convert {{ExternalCatalogTable}} to {{TableSource}}
> because we could easily get {{TableSourceTable}} from {{TableSource}}.
> Because different {{TableSource}} often contains different fields to initiate
> an instance. E.g. {{CsvTableSource}} needs path, fieldName, fieldTypes,
> fieldDelim, rowDelim and so on to create a new instance ,
> {{KafkaTableSource}} needs configuration and tableName to create a new
> instance. So it's not a good idea to let Flink framework be responsible for
> translate {{ExternalCatalogTable}} to different kind of
> {{TableSourceTable}}.
> Here is one solution. Let {{TableSource}} specify a converter.
> 1. provide an Annatition named {{ExternalCatalogCompatible}}. The
> {{TableSource}} with the annotation means it is compatible with external
> catalog, that is, it could be converted to or from ExternalCatalogTable. This
> annotation specifies the tabletype and converter of the tableSource. For
> example, for {{CsvTableSource}}, it specifies the tableType is csv and
> converter class is CsvTableSourceConverter.
> {code}
> @ExternalCatalogCompatible(tableType = "csv", converter =
> classOf[CsvTableSourceConverter])
> class CsvTableSource(...) {
> ...}
> {code}
> 2. Scan all TableSources with the ExternalCatalogCompatible annotation, save
> the tableType and converter in a Map
> 3. When need to convert {{ExternalCatalogTable}} to {{TableSource}} , get the
> converter based on tableType. and let converter do convert
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
