29 releases
0.5.3 | Aug 30, 2024 |
---|---|
0.5.0 | Jun 25, 2024 |
0.4.2 | Nov 17, 2018 |
0.4.1 | Feb 14, 2018 |
0.2.7 | Nov 2, 2016 |
#25 in #hashmap
48,203 downloads per month
Used in fewer than 13 crates
235KB
4K
SLoC
ordermap
A pure-Rust hash table which preserves (in a limited sense) insertion order.
This crate implements compact map and set data-structures, where the iteration order of the keys is independent from their hash or value. It preserves insertion order in most mutating operations, and it allows lookup of entries by either hash table key or numerical index.
Note: this crate was originally what became the indexmap
crate, and
it was deprecated for a while in favor of that, but then ordermap
returned as a wrapper over indexmap
with stronger ordering properties.
Background
This was inspired by Python 3.6's new dict implementation (which remembers the insertion order and is fast to iterate, and is compact in memory).
Some of those features were translated to Rust, and some were not. The results
were ordermap
and indexmap
, hash tables that have following properties:
- Order is independent of hash function and hash values of keys.
- Fast to iterate.
- Indexed in compact space.
- Preserves insertion order as long as you don't call
.swap_remove()
or other methods that explicitly change order.- In
ordermap
, the regular.remove()
does preserve insertion order, equivalent to whatindexmap
calls.shift_remove()
.
- In
- Uses hashbrown for the inner table, just like Rust's libstd
HashMap
does.
Since its reintroduction in 0.5, ordermap
has also used its entry order for
PartialEq
and Eq
, whereas indexmap
considers the same entries in any order
to be equal for drop-in compatibility with HashMap
semantics. Using the order
is faster, and also allows ordermap
to implement PartialOrd
, Ord
, and Hash
.
Performance
OrderMap
derives a couple of performance facts directly from how it is constructed,
which is roughly:
A raw hash table of key-value indices, and a vector of key-value pairs.
- As a wrapper,
OrderMap
should maintain the same performance asIndexMap
for most operations, with the main difference being the removal strategy. - Iteration is very fast since it is on the dense key-values.
- Lookup is fast-ish because the initial 7-bit hash lookup uses SIMD, and indices are densely stored. Lookup also is slow-ish since the actual key-value pairs are stored separately. (Visible when cpu caches size is limiting.)
- In practice,
OrderMap
has been tested out as the hashmap in rustc in PR45282 and the performance was roughly on par across the whole workload. - If you want the properties of
OrderMap
, or its strongest performance points fits your workload, it might be the best hash table implementation.
Recent Changes
See RELEASES.md.
Dependencies
~0.7–1.6MB
~28K SLoC