flatcontainer

A flat container for Rust.

[dependencies]
flatcontainer = "0.1"

Example

use flatcontainer::{FlatStack, CopyOnto};
fn main() {
  let r: Result<_, u16> = Ok("abc");
  let mut c = FlatStack::default_impl::<Result<&str, u16>>();
  c.copy(&r);
  assert_eq!(r, c.get(0));
}

Details

flatcontainer is a library that provides abstractions and implementations for flattening collections of nested data structures into dense allocations. At its core, a Region trait describes how to represent data in memory, with the option to extract a lifetimed representation of the original data, which can be different from what the caller had initially.

flatcontainer decouples the write-half of a container from its storage and read interface to permit a multitude of types to be presented to the same region. For example, a region containing string-like objects and only promising access to &str can accept anything that looks like a string, i.e., String, &str and references to said types. A region's write-half should be implemented using the CopyOnto trait.

Regions permit data access through opaque indexes, which the caller is responsible for memorizing. An index grants access to the region-specific data representation, and although it might be inspected, it should not be minted or modified in any way. As far as a region is concerned, an index is an opaque type with no particular meaning attached to it, unless specified differently by the region.

Regions roughly follow two patterns: Either they fan out to other regions, or they behave as terminal nodes and explicitly contain storage. A [Result][ResultRegion] region dispatches to an ok and error region, and uses its index type to distinguish where to look up a value. A region for slices has storage to remember slices of indexes where the index can be used to look up the datum's representation in an inner region.

flatcontainer provides FlatStack, an exemplary implementation of how to implement a collection of items that supports pushing additional elements, and retrieval by offset of previously pushed elements. It can be used in many places that simply want to use a flat data representation, but it leaves potential for optimization behind. Specifically, indexes, although opaque, follow a simple structure where a more efficient storage can save memory instead of blindly writing down all values.

All region implementations should be considered examples on how to implement specific regions that are tailored to the needs of the type, and characteristics of the data encountered in practice. This crate provides a Containerized trait to let types express their suggested region, but users can select a different region, as long as it is compatible with the types to be stored. For example, a vector suggests to use a slice-based region, but a client might know that the inner type is copy, and hence better uses a sliced-based region that does not fan out to a region for the individual elements.

Safety

This crate is safe to use, and all unsafe code can be explained locally. At the moment, this is only for assuming that utf-8 data is correct, which is true by construction.

Features

The serde feature controls whether types implement support for serializing and deserializing data. Enabled by default.

Performance and design considerations

A goal of flatcontainer is to store O(n) objects in less than O(n) allocations, for example in O(log n), or if pre-sized correctly, in constant allocations. It can achieve this by laying out data densely in memory. This comes with benefits, and restrictions. It allows fast sequential access as the CPU can prefetch data, and it avoids loading data into caches that is not needed. Reducing the number of allocations limits chatter with the allocator. On the flip side, the regions are append-only, and copying a value destructs its original form. Reading owned data requires copying or cloning.

Flatcontainer's region abstraction requires the user to store indexes. Without specific knowledge about a region, it's likely the index is best stored in a vector. In many cases, we know more about the data, and can use a better approach to storing indexes.

• Regions storing slices often have an index that looks like (start, end). It is easy to observe that the previous element's end is equal to the current element's start, so it should only be stored once. This can reduce the size of the index per element from 16 to 8 bytes.
• For index values that fit in 32 bits, we only need 4 bytes in memory. We can specialize a container for indexes that uses u32 to store values smaller than 2^32, and use u64 otherwise.
• The index into regions storing constant-sized elements often looks like strided numbers, e.g., 0, 2, 4, 8, ..., which we can represent using constant memory by remembering the stride and length. We can extend this to storing a tail of elements equals to the last stride by adding another count. Such an index container uses 0 bits in the limit!
• Consult the offsets module for types specialized to storing indexes using less bits.

Flatcontainer provides some implementations of these concepts. To merge adjacent start-end pairs, wrap a region in a ConsecutiveOffsetPairs region. It stores indexes and presents outwards as a dense sequence of 0, 1, 2, ...

A CollapseSequence region remembers the index of the last element, and if a new element equals the last, it'll return the previous index again instead of storing the same element multiple times. This is limited to the direct predecessor, because otherwise storing indexes and scanning through previous elements would be too expensive.

Comparison to columnation

Flatcontainer takes several ideas from columnation. It uses the concept of regions to abstract dense allocations while still identifying individual objects. Where columnation returns owned objects that need to be treated as if they were references, flatcontainer returns an opaque index, which a region can translate to a lifetimed type. Columnation reads and writes the same type, which forces it to accept owned data and present look-alike data outwards, forcing the user to be careful when dropping data. Flatcontainer avoids these issues because indexes do not contain pointers to owned data.

Flatcontainer offers similar performance to columnation, and is faster in some situations. For storing strings, flatcontainer provides an index of (start, end) to reconstruct the original input as a reference, while columnation returns an owned string, which roughly looks like (pointer, size, capacity). Instead of 24 bytes overhead on a 64-bit CPU, we have 16 byte overhead. We can reduce the 16 byte overhead to 8 by observing that consecutive entries start at the previous' end, allowing us to reconstruct the end by looking at the next element's start.

License