7 releases
0.0.15 | Aug 28, 2024 |
---|---|
0.0.14 | Jul 23, 2024 |
#65 in Database implementations
467 downloads per month
505KB
7.5K
SLoC
VelarixDB is an LSM-based storage engine designed to significantly reduce IO amplification, resulting in better performance and durability for storage devices.
Introduction
VelarixDB: Designed to reduce IO amplification
VelarixDB is an ongoing project (not production ready) designed to optimize data movement during load times and compaction. Inspired by the WiscKey paper, WiscKey: Separating Keys from Values in SSD-conscious Storage, velarixdb aims to significantly enhance performance over traditional key-value stores.
Problem
During compaction in LevelDB or RocksDB, in the worst case, up to 10 SSTable files needs to be read, sorted and re-written since keys are not allowed to overlapp across all the sstables from Level 1 downwards. Suppose after merging SSTables in one level, the next level exceeds its threshold, compaction can cascade from Level 0 all the way to Level 6 meaning the overall write amplification can be up to 50 (ignoring the first compaction level).[ Reference -> Official LevelDB Compaction Process Docs ]. This repetitive data movement can cause significant wear on SSDs, reducing their lifespan due to the high number of write cycles. The goal is to minimize the amount of data moved during compaction, thereby reducing the amount of data re-written and extending the device's lifetime.
Solution
To address this, we focus on whether a key has been deleted or updated. Including values in the compaction process (which are often larger than keys) unnecessarily amplifies the amount of data read and written. Therefore, we store keys and values separately. Specifically, we map value offsets to the keys, represented as 32-bit integers.
This approach reduces the amount of data read, written, and moved during compaction, leading to improved performance and less wear on storage devices, particularly SSDs. By minimizing the data movement, we not only enhance the efficiency of the database but also significantly extend the lifespan of the underlying storage hardware.
Performance Benefits
According to the benchmarks presented in the WiscKey paper, implementations can outperform LevelDB and RocksDB by:
- 2.5x to 111x for database loading
- 1.6x to 14x for random lookups
Addressing major concerns
- Range Query: Since keys are separate from values, won't that affect range queries performance. Well, we now have internal parallelism in SSDs, as we fetch the keys from the LSM tree we can fetch the values in parallel from the vlog file. This benchmark from the Wisckey Paper shows how for request size ≥ 64KB, the aggregate throughput of random reads with 32 threads matches the sequential read throughput.
- More Disk IO for Reads: Since keys are now seperate from values, we have to make extra disk IO to fetch values? Yes, but since the key density now increases for each level (since we are only storing keys and value offsets in the sstable), we will most likely search fewer levels compared to LevelDB or RocksDB for thesame query. A significant portion of the LSM tree can also be cached in memory.
Designed for asynchronous runtime (unstable)
Based on the introduction and efficiency of async IO at the OS kernel level e.g io_uring for the Linux kernel, VelarixDB is designed for asynchronous runtime. In this case Tokio runtime. Tokio allows for efficient and scalable asynchronous operations, making the most of modern multi-core processors. Frankly, most OS File System does not provide async API currently but Tokio uses a thread pool to offload blocking file system operations. This means that even though the file system operations themselves are blocking at the OS level, Tokio can handle them without blocking the main async task executor. Tokio might adopt io_uring in the future. (We haven't benchmarked the async version therefore this is unstable and might be removed in future versions)
Disclaimer
Please note that velarixdb is still under development and is not yet production-ready.
NOTE
v2 is the most recent version (not experimental) and under active development, the src modules are for the experimental version
Basic Features
- Atomic
Put()
,Get()
,Delete()
, andUpdate()
operations - 100% safe & stable Rust
- Separation of keys from values, reducing the amount of data moved during compaction (i.e., reduced IO amplification)
- Garbage Collector
- Lock-free memtable with Crossbeam SkipMap (no
Mutex
) - Tokio Runtime for efficient thread management
- Bloom Filters for fast in-memory key searches
- Crash recovery using the Value Log
- Index to improve searches on Sorted String Tables (SSTs)
- Key Range to store the largest and smallest keys in an SST
- Sized Tier Compaction Strategy (STCS)
TODO
- Snapshot Isolation
- Block Cache
- Batched Writes
- Range Query
- Snappy Compression
- Value Buffer to keep values in memory and only flush in batches to reduce IO (under investigation)
- Checksum to detect data corruption
- Leveled Compaction (LCS), Time-Window Compaction (TCS), and Unified Compaction (UCS)
- Monitoring module to continuously monitor and generate reports
It is not:
- A standalone server
- A relational database
- A wide-column database: it has no notion of columns
Constraint
- Keys are limited to 65,536 bytes, and values are limited to 2^32 bytes. Larger keys and values have a bigger performance impact.
- Like any typical key-value store, keys are stored in lexicographic order. If you are storing integer keys (e.g., timeseries data), use the big-endian form to adhere to locality.
Basic usage
cargo add velarixdb
use velarixdb::db::DataStore;
# use tempfile::tempdir;
#[tokio::main]
async fn main() {
let root = tempdir().unwrap();
let path = root.path().join("velarix");
let mut store = DataStore::open("big_tech", path).await.unwrap(); // handle IO error
store.put("apple", "tim cook").await;
store.put("google", "sundar pichai").await;
store.put("nvidia", "jensen huang").await;
store.put("microsoft", "satya nadella").await;
store.put("meta", "mark zuckerberg").await;
store.put("openai", "sam altman").await;
let entry1 = store.get("apple").await.unwrap(); // Handle error
let entry2 = store.get("google").await.unwrap();
let entry3 = store.get("nvidia").await.unwrap();
let entry4 = store.get("microsoft").await.unwrap();
let entry5 = store.get("meta").await.unwrap();
let entry6 = store.get("openai").await.unwrap();
let entry7 = store.get("***not_found_key**").await.unwrap();
assert_eq!(std::str::from_utf8(&entry1.unwrap().val).unwrap(), "tim cook");
assert_eq!(std::str::from_utf8(&entry2.unwrap().val).unwrap(), "sundar pichai");
assert_eq!(std::str::from_utf8(&entry3.unwrap().val).unwrap(), "jensen huang");
assert_eq!(std::str::from_utf8(&entry4.unwrap().val).unwrap(), "satya nadella");
assert_eq!(std::str::from_utf8(&entry5.unwrap().val).unwrap(), "mark zuckerberg");
assert_eq!(std::str::from_utf8(&entry6.unwrap().val).unwrap(), "sam altman");
assert!(entry7.is_none());
// Remove an entry
store.delete("apple").await.unwrap();
// Update an entry
let success = store.update("microsoft", "elon musk").await;
assert!(success.is_ok());
}
Store JSON
use serde::{Deserialize, Serialize};
use serde_json;
use velarixdb::db::DataStore;
use tempfile::tempdir;
#[tokio::main]
async fn main() {
let root = tempdir().unwrap();
let path = root.path().join("velarix");
let mut store = DataStore::open("big_tech", path).await.unwrap(); // handle IO error
#[derive(Serialize, Deserialize)]
struct BigTech {
name: String,
rank: i32,
}
let new_entry = BigTech {
name: String::from("Google"),
rank: 50,
};
let json_string = serde_json::to_string(&new_entry).unwrap();
let res = store.put("google", json_string).await;
assert!(res.is_ok());
let entry = store.get("google").await.unwrap().unwrap();
let entry_string = std::str::from_utf8(&entry.val).unwrap();
let big_tech: BigTech = serde_json::from_str(&entry_string).unwrap();
assert_eq!(big_tech.name, new_entry.name);
assert_eq!(big_tech.rank, new_entry.rank);
}
Examples
See here for practical examples
Dependencies
~12–24MB
~369K SLoC