elek commented on a change in pull request #1411: URL: https://github.com/apache/hadoop-ozone/pull/1411#discussion_r487496850
########## File path: hadoop-hdds/docs/content/design/s3_hcfs.md ########## @@ -0,0 +1,280 @@ +--- +title: S3/Ozone Filesystem inter-op +summary: How to support both S3 and HCFS and the same time +date: 2020-09-09 +jira: HDDS-4097 +status: draft +author: Marton Elek, +--- +<!-- + Licensed 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. See accompanying LICENSE file. +--> + +# Ozone S3 vs file-system semantics + +Ozone is an object-store for Hadoop ecosystem which can be used from multiple interfaces: + + 1. From Hadoop Compatible File Systems (will be called as *HCFS* in the remaining of this document) (RPC) + 2. From S3 compatible applications (REST) + 3. From container orchestrator as mounted volume (CSI, alpha feature) + +As Ozone is an object store it stores key and values in a flat hierarchy which is enough to support S3 (2). But to support Hadoop Compatible File System (and CSI), Ozone should simulated file system hierarchy. + +There are multiple challenges when file system hierarchy is simulated by a flat namespace: + + 1. Some key patterns couldn't be easily transformed to file system path (e.g. `/a/b/../c`, `/a/b//d`, or a real key with directory path like `/b/d/`) + 2. Directory entries (which may have own properties) require special handling as file system interface requires a dir entry even if it's not created explicitly (for example if key `/a/b/c` is created `/a/b` supposed to be a visible directory entry for file system interface) + 3. Non-recursive listing of directories can be hard (Listing direct entries under `/a` should ignore all the `/a/b/...`, `/a/b/c/...` keys) + 4. Similar to listing, rename can be a costly operation as it requires to rename many keys (renaming a first level directory means a rename of all the keys with the same prefix) + +See also the [Hadoop S3A documentation](https://hadoop.apache.org/docs/current/hadoop-aws/tools/hadoop-aws/index.html#Introducing_the_Hadoop_S3A_client) which describes some of these problem when AWS S3 is used. (*Warnings* section) + +# Current status + +As of today *Ozone Manager* has two different interfaces (both are defined in `OmClientProtocol.proto`): + + 1. object store related functions (like *CreateKey*, *LookupKey*, *DeleteKey*,...) + 2. file system related functions (like *CreateFile*, *LookupFile*,...) + +File system related functions uses the same flat hierarchy under the hood but includes additional functionalities. For example the `createFile` call creates all the intermediate directories for a specific key (create file `/a/b/c` will create `/a/b` and `/a` entries in the key space) + +Today, if a key is created from the S3 interface can cause exceptions if the intermediate directories are checked from HCFS: + + +```shell +$ aws s3api put-object --endpoint http://localhost:9878 --bucket bucket1 --key /a/b/c/d + +$ ozone fs -ls o3fs://bucket1.s3v/a/ +ls: `o3fs://bucket1.s3v/a/': No such file or directory +``` + +This problem is reported in [HDDS-3955](https://issues.apache.org/jira/browse/HDDS-3955), where a new configuration key is introduced (`ozone.om.enable.filesystem.paths`). If this is enabled, intermediate directories are created even if the object store interface is used. + +This configuration is turned off by default, which means that S3 and HCFS couldn't be used together. + +To solve the performance problems of the directory listing / rename, [HDDS-2939](https://issues.apache.org/jira/browse/HDDS-2939) is created, which propose to use a new prefix table to store the "directory" entries (=prefixes). + +[HDDS-4097](https://issues.apache.org/jira/browse/HDDS-4097) is created to normalize the key names based on file-system semantics if `ozone.om.enable.filesystem.paths` is enabled. But please note that `ozone.om.enable.filesystem.paths` should always be turned on if S3 and HCFS are both used. It means that if both S3 and HCFS are used, normalization is forced, and S3 interface is not fully AWS S3 compatible. There is no option to use HCFS and S3 but with full AWS compatibility (and reduced HCFS compatibility). + +# Goals + + * Out of the box Ozone should support both S3 and HCFS interfaces without any settings. (It's possible only for the regular, fs compatible key names) + * As 100% compatibility couldn't be achieved on both side we need a configuration to set the expectations for incompatible key names + * Default behavior of `o3fs` and `ofs` should be as close to `s3a` as possible (when s3 compatibilty is prefered) + +# Possible cases to support + +There are two main aspects of supporting both `ofs/o3fs` and `s3` together: + + 1. `ofs/o3fs` require to create intermediate directory entries (for example `/a/b` for the key `/b/c/c`) + 2. Special file-system incompatible key names require special attention + +The second couldn't be done with compromise. + + 1. We either support all key names (including non fs compatible key names), which means `ofs/o3fs` can provide only a partial view + 2. Or we can normalize the key names to be fs compatible (which makes it possible to create inconsistent S3 keys) + +HDDS-3955 introduced `ozone.om.enable.filesystem.paths`, with this setting we will have two possible usage pattern: + +| ozone.om.enable.filesystem.paths= | true | false +|-|-|-| +| create itermediate dirs | YES | NO | +| normalize key names from `ofs/o3fs` | YES | NO +| force to normalize key names of `s3` interface | YES | NO +| `s3` key `/a/b/c` available from `ofs/o3fs` | YES | NO +| `s3` key `/a/b//c` available from `ofs/o3fs` | YES | NO +| `s3` key `/a/b//c` available from `s3` | AWS S3 incompatibility | YES Review comment: In general, AWS S3 can save any key without normalization, and the same content can be retrieved with the same key (and list shows the raw names). `AWS S3 incompatibility`: means partial incompatibility. All the `s3` tools which depends on the original AWS S3 behavior can be failed. (For example if a S3 Fuse file system creates dirs with keys ending with `/` **AND** store real directory metadata on that specific key --> it will be broken). I started to create robot tests for these cases which can be used to check this side of the compatibility: https://github.com/elek/hadoop-ozone/commit/d21890bb0b0d08d0f8c631e9f24af4745b2f3aca ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: ozone-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: ozone-issues-h...@hadoop.apache.org
