anilkumgopalap commented on code in PR #14862: URL: https://github.com/apache/gravitino/pull/14862#discussion_r4603202750
########## catalogs/catalog-lakehouse-generic/src/main/java/org/apache/gravitino/catalog/lakehouse/generic/SupportsSchemaLocations.java: ########## @@ -0,0 +1,111 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package org.apache.gravitino.catalog.lakehouse.generic; + +import javax.annotation.Nullable; + +/** + * An optional capability of a {@link TableLocationProvider}: allocating the location of a schema + * when it is created, and releasing it when the schema is dropped. + * + * <p>A provider opts in by implementing this interface alongside {@link TableLocationProvider}. + * Services that allocate storage usually do so hierarchically -- a schema is registered first, with + * the owner and the storage domain it belongs to, and the paths of its tables are derived from it + * -- and the table callbacks alone offer no point at which to register the schema. + * + * <p>A provider that does not implement it is never consulted for schemas, and the catalog does no + * extra work on its behalf: dropping an empty schema still does not read the schema. That is why + * this is a separate interface rather than default methods on {@link TableLocationProvider}. + * + * <p>The division of responsibilities is the one {@link TableLocationProvider} describes. The + * provider owns what it brought into being so that the schema location can be used -- its + * registration, its quota, any container it created -- and the table formats own the data written + * under it. Like the table callbacks, both methods here are called concurrently and must be + * thread-safe, and the catalog applies no timeout to them. + * + * <p><b>Known limitations.</b> + * + * <ul> + * <li>{@link #provisionSchemaLocation(SchemaLocationContext)} is called before the schema is + * stored. A creation that fails afterwards -- a schema of the same name created concurrently, + * or a store failure -- leaves a location provisioned for a schema that does not exist, and + * there is no compensating unprovision. A provider that answers idempotently may have + * returned the allocation of the schema that won the race, and releasing it would take it + * from a schema that exists. A schema that already exists is detected before the provider is + * consulted, so the cases that remain are the races. + * <li>{@code location} is a schema property, and altering it does not tell the provider. + * <li>{@link #unprovisionSchemaLocation(SchemaLocationContext)} is called before the schema is + * removed. A removal that fails afterwards, or a server that stops in between, leaves the + * schema in place with its location already released; dropping it again calls the provider + * again for the same schema. + * </ul> + */ +public interface SupportsSchemaLocations { + + /** + * Provisions the location of the schema that is being created. + * + * <p>The context carries the properties of the creation request, including a caller-supplied + * {@code location}, if any. As with tables, what happens to a supplied location is the provider's + * decision: return null to keep it, return something else to place the schema elsewhere, or throw + * to refuse the creation. + * + * <p>A non-null result is stored verbatim as the schema's {@code location} property. It must not + * be blank; the catalog rejects the creation with an {@link IllegalArgumentException} otherwise. + * Because the built-in provider derives the location of a table from its schema's {@code + * location}, recording it also keeps the schema consistent with the tables placed under it. + * + * <p>If this method throws, the schema is not created. + * + * @param context the schema being created, with the properties of the creation request + * @return the location to store on the schema, or null to keep the properties of the request + * unchanged + */ + @Nullable + String provisionSchemaLocation(SchemaLocationContext context); + + /** + * Releases what this provider allocated for a schema that is being dropped. + * + * <p>It is called after the tables of the schema are dropped and before the schema itself is + * removed. With cascade, the tables go through the same path as a single table drop, so their + * {@link TableLocationProvider#unprovisionTableLocation(TableLocationContext)} callbacks have + * already run. The context carries the stored properties of the schema and which of the cascaded + * tables kept their data. + * + * <p><b>The catalog reports; the provider decides.</b> A table that kept its data is an external + * table that was dropped rather than purged: the table format left its dataset in place and its + * location was deliberately not unprovisioned. When the context lists any, the schema location + * may still hold data, and the provider decides whether to release it anyway, keep it, or keep it + * and report it. There is deliberately no force option on the drop: the decision belongs to the + * provider, as it does for tables. + * + * <p>What the provider must not touch is the same as for tables: the data. It may remove a + * container it created itself only once that container holds nothing but what the provider put + * there. + * + * <p>If this method throws, the drop fails and the schema stays in place, so the drop can be + * retried; tables the cascade already dropped stay dropped. Because the schema is removed only + * after this method returns, a retry can call it again for a schema whose location it already + * released, so it must be idempotent. + * + * @param context the schema being dropped and the tables its cascade dropped without their data + */ + void unprovisionSchemaLocation(SchemaLocationContext context); Review Comment: Good idea, done, and I applied it to `unprovisionTableLocation` too so the two stay consistent. Both return `true` if the provider released something and `false` if there was nothing to release (e.g. already released on an earlier attempt). The schema default and `DefaultLocationProvider` return `false`. The catalog logs a `false` result at DEBUG and doesn't otherwise act on it, so the outcome of the drop is unchanged; that leaves room to use it later without another SPI change. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
