content aware chunking

未关闭
#508 1 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
25/100
Issue 类型
功能
描述清晰度
需要澄清
活跃度
停滞
技术栈
go
领域
backend

调研方向

未指定文件或测试。首先定位 UnixFS chunker 和 decoder 的入口点,然后确定所提议的格式感知行为的范围;当无效输入仍能通过现有 decoder 按字节往返,并且生成的文件保留预期的去重链接时,即视为完成。

由索引模型根据 Issue 内容生成。

描述

kind/enhancement P3

One hard thing with deduplication over WORM merkle-tree datastructures like unixfs is that we are extremely limited when finding backrefs. Usually deduplication code is able to search for backrefs freely and fixup the old and new datastructure so that they point to a shared piece of data.
Here we can't, consider this case:

# file 1
| chunk1       |
| aaaabbbbcccc | 
# file 2
| chunk1 |
| bbbb   | 

Assuming file 1 already exists we can't do any deduplication when chunking file 2.
We would have needed to chunk file 1 this way originally:

# file 1
|  c1  |  c2  |  c3  |
| aaaa | bbbb | cccc | 

However when we are chunking file 1 we don't already know what file 2's structure is.


It would be nice to have a chunker which use applicative usecase to do so.
We could try to match known file formats and change strategy based on the file.
We should apply this logic recursively based on contextual hints.

Various strategy I can think off (mainly container formats):

  • tar, cpio, iso, overlay2fs (used by docker), chunk each file recursively, embed headers in the root blocks.
    Deduplicate files inside and outside the archive. As well as deduplicate partial updates where only some files in the archive changed.
  • car, chunk each block using raw leaves, embed headers in the root blocks.
    Deduplicate blocks inside and outside the car. As well as support partial updates where only some blocks in the archive changed. Pinning a car should be the same as pinning the content of this car plus some metadata pixie dust.
  • video containers, chunk streams and sections recursively.
    Deduplicate alternative tracks like audio or captions, deduplicate across splices (*if the container allow low level concatenation)
    Maybe would require a custom video encoder which takes care of reusing bytes as-is in the most amount of possible cases to be really effective.
  • video codecs, chunk around keyframe and delta
    Avoid read amplifications when seeking.
  • gzip, zstd, (other low level concatenable compression streams), chunk around sections
    Formats like stargz compress each file within an archive individually and then concatenate all the resulting streams. This works because for theses decompress(compress(a) + compress(b)) is equal to decompress(compress(a + b)).
    Deduplicate compressed files inside and outside the archive, as well as partial updates where only some files changed.
  • zip, jar, chunk compressed sections individually
    Deduplicate partial updates where only some files in the archive changed.
  • ... ideas welcome

The chunker would need to be resistant and tolerant, if a file is incorrectly classified or does not follow the underlying format correctly. The chunker must still produce a valid unixfs file which once deserialized gives byte for byte identical copy of the original input.
In other words, if someone tries to add an invalid tar file, the chunker must create a unixfs file which encode this invalid tar file.


This solution is simpler than allowing arbitrary backrefs in other blocks because it ensure files don't grow by adding more backrefs. We don't want to find ourself in the situation where a file is bigger to download from scratch than a non de-duped alternative because we assumed that a bunch of blocks would already be cached.
It also does not require synchronizing matches, we can have deduplication without ever comparing old and new versions.


In this case we are not only considering where are leaves are placed but also how theses are linked together.
If I add a tar it's not good enough that it use the same leaves as the file outside of the tar, we want pinning the tar to be equivalent to also pin the file outside of the tar, so the root cid of the file must be linked from the tar's unixfs dag.


The resulting files would be compatible with existing unixfs decoder and not require out-of-band file format knowledge, this is purely an improvement in how we convert from non unixfs to unixfs.
The goal is not to add a webm or tar codec which everyone would need to be updated to understand.

主要语言
Go
星标
326
派生
165
PR 合并指标
30 天内没有已合并 PR

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 有 Pull Request 模板
  • 没有贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

ipfs/boxo 的其他 Issue

查看 ipfs/boxo 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。