about summary refs log tree commit diff
path: root/tvix/castore/docs
diff options
context:
space:
mode:
authorFlorian Klink <flokli@flokli.de>2023-11-02T08·33+0200
committerclbot <clbot@tvl.fyi>2023-11-02T09·08+0000
commit5a0ae8312c58fb67371c3c5961a39b9a5f9c87e6 (patch)
tree802d61e9f3409c497397f56a63e259504248315e /tvix/castore/docs
parentbeae3a4bf163c34b90530ab4601ce4dc753ed9e4 (diff)
docs(tvix/castore): add notes on verified streaming r/6921
These are still a bit WIP, but should already explain the idea better
than not having any documentation about how we intend to implement this
at all.

Let's put this into the repo, as a working document.

Co-Authored-By: edef <edef@edef.eu>
Change-Id: Ic0c5195cdf0a33c8382af1b5cff9064fe6ef9ffb
Reviewed-on: https://cl.tvl.fyi/c/depot/+/9894
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
Diffstat (limited to 'tvix/castore/docs')
-rw-r--r--tvix/castore/docs/verified-streaming.md121
1 files changed, 121 insertions, 0 deletions
diff --git a/tvix/castore/docs/verified-streaming.md b/tvix/castore/docs/verified-streaming.md
new file mode 100644
index 000000000000..24197d90a013
--- /dev/null
+++ b/tvix/castore/docs/verified-streaming.md
@@ -0,0 +1,121 @@
+# Verified streaming
+
+`//tvix/castore` is a content-addressed storage system, using [blake3] as hash
+function.
+
+This means returned data is fetched by using the digest as lookup key, and can
+be verified to be correct by feeding the received data through the hash function
+and ensuring it matches the digest initially used for the lookup.
+
+This means, data can be downloaded by any untrusted third-party as well, as the
+received data is validated to match the digest it was originally requested with.
+
+However, for larger blobs of data, having to download the entire blob to be able
+to determine whether it's correct before being able to return it to an upper
+layer takes a lot of time, and is wasteful, if we're only interested in a small
+portion of it.
+
+Especially when substituting from an untrusted third-party, we want to be able
+to detect quickly if that third-party is sending us wrong data, and terminate
+the connection early.
+
+## Chunking
+
+This problem has historically been solved by exchanging a list of smaller
+chunks, which can be fetched individually.
+
+BitTorrent for example breaks files up into smaller chunks, and maintains a list
+of sha1 digests for each of these chunks. After the list has been fetched, this
+allows fetching smaller parts of data selectively from untrusted third-parties.
+
+Similarly, IPFS uses its IPLD model to content-address a Merkle DAG of chunk
+nodes.
+
+While these approaches solve the problem of being able to fetch smaller chunks,
+they have a big disadvantage: the chunking parameters, and the topology of
+the graph structure itself "bleed" into the hash of the entire data structure
+itself.
+
+This comes with some disadvantages:
+
+Depending on the chunking parameters used, there's different representations for
+the same data, causing less data sharing/reuse in the overall content- addressed
+system, both when downloading data from third-parties, as well as benefiting
+from data already available locally.
+
+This can be workarounded by agreeing on only single way of chunking, but it's
+not pretty.
+
+## Chunking in tvix-castore
+
+tvix-castore uses BLAKE3 as a digest function, which internally uses a fixed
+chunksize of 1024 bytes.
+
+BLAKE3 is a tree hash where all left nodes fully populated, contrary to
+conventional serial hash functions. To be able to validate the hash of a node,
+one only needs the hash of the (2) children, if any.
+
+This means one only needs to the root digest to validate a construction, and
+lower levels of the tree can be omitted.
+
+This relieves us from the need of having to communicate more granular chunking
+upfront, and making it part of our data model.
+
+## Logical vs. physical chunking
+
+Due to the properties of the BLAKE3 hash function, we have logical blocks of
+1KiB, but this doesn't necessarily imply we need to restrict ourselves to these
+chunk sizes.
+
+The only thing we need to be able to read and verify an arbitrary byte range is
+having the covering range of aligned 1K blocks.
+
+## Actual implementation
+
+ -> BlobService.Read() gets the capability to read chunks as well
+ -> BlobService.Stat() can hand out a list of chunks with a given chunksize if asked so.
+      rq params: send_bao bool
+         server should be able to offer bao all the way down to 1k
+         some open questions w.r.t sending the whole bao until there, or just
+         all the hashes on the "most granular" level
+         -> we can recreate everything above up to the root hash.
+         -> can we maybe add this to n0-computer/bao-tree as another outboard format?
+      resp:
+        - bao_shift: how many levels on the bottom were skipped.
+          0 means send all the leaf node hashes (1K block size)
+        - "our bao": blake3 digests for a given static chunk size + path down to the last leaf node and its data (length proof)
+        - list of (b3digest,size) of all physical chunks.
+          The server can do some CDC on ingestion, and communicate these chunks here.
+          Chunk sizes should be a "reasonable size", TBD, probably something between 512K-4M
+
+Depending on the bao depth received from the server, we end up with a logical
+size of chunks that can be fetched in an authenticated fashion.
+
+Assuming the bao chunk size received is 4(*1KiB bytes) (`bao_shift` of 2), and a
+total blob size of 14 (*1KiB bytes), we can fetch
+`[0..=3]`, `[4..=7]`, `[8..=11]` and `[12..=13]` in an authenticated fashion:
+
+`[ 0 1 2 3 ] [ 4 5 6 7 ] [ 8 9 10 11 ] [ 12 13 ]`
+
+Assuming the server now informs us about the following physical chunking:
+
+`[ 0 1 ] [ 2 3 4 5 ] [ 6 ] [ 7 8 ] [ 9 10 11 12 13 14 15 ]`
+
+To read from 0 until 4 (inclusive), we need to fetch physical chunks
+`[ 0 1 ]`, `[ 2 3 4 5 ]` and `[ 6 ] [ 7 8 ]`.
+
+`[ 0 1 ]` and `[ 2 3 4 5 ]` are obvious, they contain the data we're
+interested in.
+
+We however also need to fetch the physical chunks `[ 6 ]` and `[ 7 8 ]`, so we
+can assemble `[ 4 5 6 7 ]` to verify that logical chunk.
+
+Each physical chunk fetched can be validated to have the blake3 digest that was
+communicated upfront, and can be stored in a client-side cache/storage.
+
+If it's not there, the client can use the `BlobService.Read()` interface with
+the *physical chunk digest*.
+
+---
+
+[blake3]: https://github.com/BLAKE3-team/BLAKE3
\ No newline at end of file