Key-value database for embedded systems, for raw NOR flash, using an LSM-Tree.
crates.io - Docs - Docs for git - Chat
- Keys and values are arbitrary-length byte arrays. The database is essentially an on-disk
Map<Vec<u8>, Vec<u8>>
. O(log n)
reads. AmortizedO(log n)
writes. No linear scans!- Power-fail safe. Powering off in the middle of writes never corrupts the database.
- Transaction support:
- Atomic writes: Start a write transaction, write multiple keys, commit. If power fails midway, either all or no writes are committed.
- Consistent reads: Read transactions see a consistent snapshot of the database, unaffected by concurrent writes.
- Unlimited read transactions and one write transaction are allowed concurrently.
- Read transactions are only blocked by a write transaction commit, not by the whole write transaction. Commit is fast,
O(1)
.
- Wear leveling: erase cycles are spread out evenly between all flash pages. Pages are allocated cyclically. At boot, a random seed is required to decide which is the first.
- Corruption-resistant: A corrupted or deliberately manipulated flash image cannot cause crashes, panics or infinite loops, only
Err(Corrupted)
errors. - Optional CRC32 protection of headers and data on flash.
- Extensively tested, using unit tests and fuzzing.
The project is in production ready stage. The author is using it in production, it is fully functional and has no known bugs. This does not mean it's feature-complete, see the limitations below.
The on-disk format is not stable yet.
- Major releases can do fully breaking changes to the on-disk format. New code won't be able to read old databases and vice-versa. No code to upgrade in-place will be provided. You'll have to either format and lose all data, or read out the data and write it back in the newer format.
- Minor releases can do backwards-compatible changes: they'll always be able to read databases written by older versions, but they might write databases using new features that older versions can't read.
- Patch releases will maintain full backwards and forwards on-disk compatibility.
- Optimize tiny write transactions: append to the last file if possible, instead of starting a new one. Currently each write transaction opens a new file, which will have to erase at least one full page, even if the transaction writes just one small key. It is recommended to batch multiple writes in a single transaction for performance.
- Support access align higher than 4. Currently reads/writes are (optionally) aligned up to 4 bytes. Some flash out there can only be written in 8-byte words or higher.
- Add a max chunk size, to reduce the RAM requirement in PageReader.
- Allow writes within a transaction to be unsorted.
- Allow reads within a write transaction. They should see the the not yet committed writes in the current transaction.
- Allow iterating the records in the database.
- Add optional encryption + authentication support (which disables CRCs)
- Integrate with
embedded-storage
.
ekv
works best for datasets with large amounts of keys (>1000), where its O(log n)
complexity outperforms linear search. For datasets with less keys, other key-value databases based on linear search (such as sequential-storage
) will be faster.
- If the dataset fits in RAM, you could read/write it all at once. Either making it
repr(C)
and transmuting, or serializing it with some compactserde
flavor such aspostcard
- sequential-storage - Rust. Linear search. No multi-key transactions. Multiple single-key transactions can be written to the same page, unlike
ekv
. - Tock's tickv - Rust. Hash table. No multi-key transactions.
- pigweed's pw_kvs - C. Linear search. No multi-key transactions. Multiple single-key transactions can be written to the same page, unlike
ekv
.
Embedded Key-Value! :)
This work is licensed under either of
- Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.