[GitHub] [hadoop] ahmarsuhail commented on a diff in pull request #4205: HADOOP-18177. Document prefetching architecture.

[GitBox <gi...@apache.org>]

ahmarsuhail commented on code in PR #4205:
URL: https://github.com/apache/hadoop/pull/4205#discussion_r857842464


##########
hadoop-tools/hadoop-aws/src/site/markdown/tools/hadoop-aws/prefetching.md:
##########
@@ -0,0 +1,151 @@
+<!---
+  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.
+-->
+
+# S3A Prefetching
+
+
+This document explains the `S3PrefetchingInputStream` and the various components it uses.
+
+This input stream implements prefetching and caching to improve read performance of the input stream. A high level overview of this feature can also be found on [this](https://medium.com/pinterest-engineering/improving-efficiency-and-reducing-runtime-using-s3-read-optimization-b31da4b60fa0) blogpost.
+
+With prefetching, we divide the file into blocks of a fixed size (default is 8MB), associate buffers to these blocks, and then read data into these buffers asynchronously. We also potentially cache these blocks.
+
+### Basic Concepts
+
+* **File** : A binary blob of data stored on some storage device.
+* **Block :** A file is divided into a number of blocks. The default size of a block is 8MB, but can be configured. The size of the first n-1 blocks is same,  and the size of the last block may be same or smaller.
+* **Block based reading** : The granularity of read is one block. That is, we read an entire block and return or none at all. Multiple blocks may be read in parallel.
+
+### Configuring the stream
+
+|Property    |Meaning    |Default    |
+|---	|---	|---	|
+|fs.s3a.prefetch.enabled    |Enable the prefetch input stream    |TRUE |
+|fs.s3a.prefetch.block.size    |Size of a block    |8MB    |
+|fs.s3a.prefetch.block.count    |Number of blocks to prefetch    |8    |
+
+### Key Components:
+
+`S3PrefetchingInputStream` - When prefetching is enabled, S3AFileSystem will return an instance of this class as the input stream. Depending on the file size, it will either use the `S3InMemoryInputStream` or the `S3CachingInputStream` as the underlying input stream.
+
+`S3InMemoryInputStream` - Underlying input stream used when the file size < configured block size. Will read the entire file into memory.
+
+`S3CachingInputStream` - Underlying input stream used when file size > configured block size. Uses asynchronous prefetching of blocks and caching to improve performance.
+
+`BlockData` - Holds information about the blocks in a file, such as:
+
+* Number of blocks in the file
+* Block size
+* State of each block (initially all blocks have state *NOT_READY*). Other states are: Queued, Ready, Cached.
+
+`BufferData` - Holds the buffer and additional information about it such as:
+
+* The block number this buffer is for
+* State of the buffer (Unknown, Blank, Prefetching, Caching, Ready, Done). Initial state of a buffer is blank.
+
+`CachingBlockManager` - Implements reading data into the buffer, prefetching and caching.
+
+`BufferPool` - Manages a fixed sized pool of buffers. It’s used by `CachingBlockManager` to acquire buffers.
+
+`S3File` - Implements operations to interact with S3 such as opening and closing the input stream to the S3 file.
+
+`S3Reader` - Implements reading from the stream opened by `S3File`. Reads from this input stream in blocks of 64KB.
+
+`FilePosition` - Provides functionality related to tracking the position in the file. Also gives access to the current buffer in use.
+
+`SingleFilePerBlockCache` - Responsible for caching blocks to the local file system. Each cache block is stored on the local disk as a separate file.
+
+### Operation
+
+### S3InMemoryInputStream
+
+If we have a file with size 5MB, and block size = 8MB. Since file size is less than the block size, the `S3InMemoryInputStream` will be used.
+
+If the caller makes the following read calls:
+
+
+```
+in.read(buffer, 0, 3MB);
+in.read(buffer, 0, 2MB);
+```
+
+When the first read is issued, there is no buffer in use yet. We get the data in this file by calling the `ensureCurrentBuffer()` method, which ensures that a buffer with data is available to be read from.
+
+The `ensureCurrentBuffer()` then:
+
+* Reads data into a buffer by calling `S3Reader.read(ByteBuffer buffer, long offset, int size)`
+*  `S3Reader` uses `S3File` to open an input stream to the S3 file by making a `getObject()` request with range as `(0, filesize)`.
+*  The S3Reader reads the entire file into the provided buffer, and once reading is complete closes the S3 stream and frees all underlying resources.
+* Now the entire file is in a buffer, set this data in `FilePosition` so it can be accessed by the input stream.
+
+The read operation now just gets the required bytes from the buffer in `FilePosition`.
+
+When the second read is issued, there is already a valid buffer which can be used. Don’t do anything else, just read the required bytes from this buffer.
+
+### S3CachingInputStream
+
+
+
+[Image: image.png]

Review Comment:
   sorry, I didn't mean to commit this. I was initially using the image in the blogpost: https://medium.com/pinterest-engineering/improving-efficiency-and-reducing-runtime-using-s3-read-optimization-b31da4b60fa0, but that diagram is very high level so I'm not sure if it adds much here. Do we think having that will help? Or do we think we need a lower level architecture diagram? 



-- 
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: common-issues-unsubscribe@hadoop.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: common-issues-help@hadoop.apache.org