# Erasure Coding Layouts

## Sequential Layout (AKA Horizontal)(Current Codex implementation)

This layout is called **Sequential** because sequential blocks get stored in the same node.

### Pros
* Easy and natural implementation
* Disk are usually optimized for sequential reads so this can give a boost in performance

~~Sequential blocks get distributed across multiple nodes, which might increase the download speed as one can simultaneously download in parallel from multiple nodes~~

### Cons
* Sequential blocks get stored on the same node. One a node is trying to read just a part of the dataset (e.g., blocks 1, 2, and 3) they have to come from the same node (Node 1) which could hurt download performance as all the blocks come the same node.
* Parity data is accumulated on only a few nodes (Nodes 5 and 6)


## Rotated Sequential Layout (AKA Rotated Horizontal)

> Leo: I agree with Bulat, there is no need to show this one at this point


## Striped Layout (AKA Vertical)

This layout is called **Striped** because data gets stored by stripes across the nodes, instead of sequentially in the same node.

### Pros
* Easy and natural implementation
* Disks are usually optimized for sequential reads so this can give a boost in performance
* Sequential blocks get distributed across multiple nodes, which might increase the download speed as one can simultaneously download in parallel from multiple nodes

### Cons
* Parity data accumulated on a few nodes


## Rotated Striped Layout (AKA Rotated Vertical)

This layout is called **Rotated Striped** because in addition to the data being stored in stripes, the data is rotated progressively to get an homogeneous distribution of original and parity data across the nodes.

### Pros

* Since the data is rotated progressively, all nodes get the same ratio of download requests, as opposed to the non-rotated case where nodes 1, 2, 3, and 4 will usually have to provide more bandwidth for downloads than nodes 5 and 6, because the parity data is mostly required for repairs

### Cons

> Leo: I don't see many cons about this one


## Diagonal Layout

> Leo: I really find this layout very weird and I don't think it provides any benefits from the previous one, Rotated Stripes, so I don't even know if it is worth mentioning it.


# Erasure Coding

We present an interleaved erasure code that is able to encode arbitrary amounts of data despite the relatively small Reed-Solomon codeword size.

## Overview

Erasure coding has several complementing uses in the Codex system. It is superior to replication because it provides greater redundancy at lower storage overheads; it strengthens our remote storage auditing scheme by allowing to check only a percentage of the blocks rather than checking the entire dataset; and it helps with downloads as it allows retrieving $K$ arbitrary pieces from any nodes, thus avoiding the so called "stragglers problem".

We employ an Reed-Solomon code with configurable $K$ and $M$ parameters per dataset and an interleaving structure that allows us to overcome the limitation of a small Galois field (GF) size.

## Limitations of small GF fields

### Terminology

- A $codeword$ ($C$) is the maximum amount of data that can be encoded together, it cannot surpass the size of the GF field. For example $C <= GF(2^8) = C <= 256$.
- A $symbol$ ($s$) is an element of the GF field. A $codeword$ is composed of a number of symbols up to the size of the GF field. For example a field of $GF(2^8) = 256$ contains up to 256 symbols. The size of the symbol in bytes is also limited by the size of the GF field, a symbol in $GF(2^8)$ will have a max size of 8 bits or 1 byte.

The size of the codeword determines the maximum number of symbols that an error correcting code can encode and decode as a unit. For example, a Reed-Solomon code that uses a Galois field of $2^8$ can only encode and decode 256 symbols together. In other words, the size of the field imposes a natural limit on the size of the data that can be coded together.

Why not use a larger field? The limitations are mostly practical, either memory or CPU bound, or both. For example, most implementation rely on one or more arithmetic tables which have a linear storage overhead proportional to the size of the field. In other words, a fields of $2^{32}$ would require generating and storing several 4GB tables and it would still only allow coding 4GB of data at a time, clearly this isn't enough when the average high definition video file is several times larger than that and specially when the expectation is to allow handling very large, potentially terabyte size datasets, routinely employed in science and big-data.

## Interleaving

Interleaving is the process of combining or interweaving several symbols from disparate parts of the source data in ways that minimizes the likelihood that any sequence of these symbols damaged together compromises the rest of the data.

In our case, the primary requirement is to ensure that we're preserving the dataset in its entirety and any $K$ elements are still enough to recover the original dataset. This is important to emphasize, we cannot simply encode a chunk of data individually, that would only protect that individual chunk, the code should protect the dataset in its entirety, even if it is terabytes size.

A secondary but important requirement is to allow appending data without having to re-encode the entire dataset.

### Codex Interleaving

Codex employs a type of interleaving where every $K$ symbols spaced at $S$ intervals, are encoded with additional $M$ parity symbols equally spaced at $S$ intervals. This results in a logical matrix where each row is $S$ symbols long and each column is $N=K+M$ symbols tall.

The algorithm looks like the following:

1) Given a dataset chunked in equally sized chunks and rounded up to the next multiple of $K$, where $K$ represents the number of symbols to be coded together and $M$ the resulting parity symbols.
2) Take $K$, $S$ spaced chunks and $K$ symbols $s$ from each of the selected chunks at offset 0 and encode into $M$ parity symbols, each placed in new $S$ spaced chunks at the same offset and appended to the end of the sequence of original chunks. Repeat this steps at offset $1*sizeof(GF(p))$ and so on.
3) Repeat the above steps for the length of the entire dataset.

Bellow is a graphical outline of the process:

The sequence of original blocks


The logical matrix resulting from stacking each $S$ symbols together, where $K=4$ and $S=3$.


The resulting matrix with parity chunks added, where $K=4$, $M=2$, $N=6$ and $S=3$. Each chunk is composed by multiple EC symbols $s$ and each column of symbols are coded together.


The resulting structure is a matrix of height $S$ and width $N$. It allows loosing any $M$ "columns" of the matrix and still reconstructing the original dataset. Of course, loosing more than $M$ symbols in a given row would still render the entire dataset invalid, but the likelihood of that happening is mitigated by placing the individual original and parity chunks on independent nodes. The collection of all the chunks stored on a the same node is called an EC block.


Moreover, the resulting dataset is still systematic, which means that the original blocks are unchanged and can be used without prior decoding.

## Data placement and erasures

The code introduced in the above section satisfies our original requirement of any $K$ blocks allowing to reconstruct the original dataset. In fact, one can easily see that every row in the resulting matrix is a standalone element and any $K$ rows will allow reconstructing the original dataset.

However, the code is $K$ rows strong only if we operate under the assumption that each failure is independent of each other. Thus, it is a requirement that each row is placed independently. Moreover, the overall strength of the code decreases based on the number of dependent failures. If we place $N$ rows on two independent locations, we can only tolerate $M=N/2$ failures, three will allow tolerating $M=N/3$ failures and so on. Hence, the code is only as strong as the number of independent locations each element is stored on. Thus, each row and better yet, each element (block) of the matrix should be stored independently and in a pattern mitigating dependence, meaning that placing elements of the same column together should be avoided.

### Load balancing retrieval

There is an issue resulting mostly from the systematic nature of the code and the presence of original and parity data.

It is safe to assume that the systematic (original) data is going to be retrieved more frequently than the parity data, which should only be retrieved when some of the systematic pieces have gone missing thus, in order to prevent congestion and overloading of the systematic nodes some further placement considerations should be taken in to account. Namely, having nodes dedicated to parity data should be avoided.

This can be accomplished by choosing a placement pattern that avoids favoring some nodes over others when it comes to storing the systematic data, and at the same time it should not break the placement rules around the safety of the code. For example, one can think of a placement strategy where the node gets assigned a block based on the node's position in some queue or its id or a combination thereof.

Some placing strategies will be explored in a further document.

### Adversarial vs random erasures

There is still a problem that needs to be addressed - adversarial erasures. Both the code and the placement rules protect against random erasures, meaning erasures that aren't targeted or coordinated in any way, it doesn't protect against adversarial erasures.

In codex, adversarial erasures are partly addressed by the presence of incentives and penalties. Here we make all the standard assumptions around rational behavior, and rely on remote auditing to detect faults, penalize nodes and rebuilt the lost data when faults are detected.