A Key/Value Store For SSDs
Contents
Introduction
NuDB is an append-only, key/value store specifically optimized for random read performance on modern SSDs or equivalent high-IOPS devices. The most common application for NuDB is content addressable storage where a cryptographic digest of the data is used as the key. The read performance and memory usage are independent of the size of the database. These are some other features:
- Low memory footprint
- Database size up to 281TB
- All keys are the same size
- Append-only, no update or delete
- Value sizes from 1 to 2^32 bytes (4GB)
- Performance independent of growth
- Optimized for concurrent fetch
- Key file can be rebuilt if needed
- Inserts are atomic and consistent
- Data file may be efficiently iterated
- Key and data files may be on different devices
- Hardened against algorithmic complexity attacks
- Header-only, no separate library to build
Description
This software is close to final. Interfaces are stable. For recent changes see the CHANGELOG.
NuDB has been in use for over a year on production servers running rippled, with database sizes over 3 terabytes.
Requirements
- Boost 1.58 or higher
- C++11 or greater
- SSD drive, or equivalent device with high IOPS
These components are optionally required in order to build the tests and examples:
- CMake 3.7.2 or later (optional)
- Properly configured bjam/b2 (optional)
Example
This complete program creates a database, opens the database, inserts several key/value pairs, fetches the key/value pairs, closes the database, then erases the database files. Source code for this program is located in the examples directory.
#include <nudb/nudb.hpp>
#include <cstddef>
#include <cstdint>
int main()
{
using namespace nudb;
std::size_t constexpr N = 1000;
using key_type = std::uint32_t;
error_code ec;
auto const dat_path = "db.dat";
auto const key_path = "db.key";
auto const log_path = "db.log";
create<xxhasher>(
dat_path, key_path, log_path,
1,
make_salt(),
sizeof(key_type),
block_size("."),
0.5f,
ec);
store db;
db.open(dat_path, key_path, log_path, ec);
char data = 0;
// Insert
for(key_type i = 0; i < N; ++i)
db.insert(&i, &data, sizeof(data), ec);
// Fetch
for(key_type i = 0; i < N; ++i)
db.fetch(&i,
[&](void const* buffer, std::size_t size)
{
// do something with buffer, size
}, ec);
db.close(ec);
erase_file(dat_path);
erase_file(key_path);
erase_file(log_path);
}Building
NuDB is header-only so there are no libraries to build. To use it in your
project, simply copy the NuDB sources to your project's source tree
(alternatively, bring NuDB into your Git repository using the
git subtree or git submodule commands). Then, edit your build scripts
to add the include/ directory to the list of paths checked by the C++
compiler when searching for includes. NuDB #include lines will look
like this:
#include <nudb/nudb.hpp>
To link your program successfully, you'll need to add the Boost.Thread and Boost.System libraries to link with. Please visit the Boost documentation for instructions on how to do this for your particular build system.
NuDB tests require Beast, and the benchmarks require RocksDB. These projects are linked to the repository using git submodules. Before building the tests or benchmarks, these commands should be issued at the root of the repository:
git submodule init
git submodule update
For the examples and tests, NuDB provides build scripts for Boost.Build (b2) and CMake. To generate build scripts using CMake, execute these commands at the root of the repository (project and solution files will be generated for Visual Studio users):
cd bin
cmake .. # for 32-bit Windows build
cd ../bin64
cmake .. # for Linux/Mac builds, OR
cmake -G"Visual Studio 14 2015 Win64" .. # for 64-bit Windows builds
To build with Boost.Build, it is necessary to have the b2 executable
in your path. And b2 needs to know how to find the Boost sources. The
easiest way to do this is make sure that the version of b2 in your path
is the one at the root of the Boost source tree, which is built when
running bootstrap.sh (or bootstrap.bat on Windows).
Once b2 is in your path, simply run b2 in the root of the Beast repository to automatically build the required Boost libraries if they are not already built, build the examples, then build and run the unit tests.
On OSX it may be necessary to pass "toolset=clang" on the b2 command line. Alternatively, this may be site in site-config.jam or user-config.jam.
The files in the repository are laid out thusly:
./
bench/ Holds the benchmark sources and scripts
bin/ Holds executables and project files
bin64/ Holds 64-bit Windows executables and project files
examples/ Holds example program source code
extras/ Additional APIs, may change
include/ Add this to your compiler includes
nudb/
test/ Unit tests and benchmarks
tools/ Holds the command line tool sources
Algorithm
Three files are used.
- The data file holds keys and values stored sequentially and size-prefixed.
- The key file holds a series of fixed-size bucket records forming an on-disk hash table.
- The log file stores bookkeeping information used to restore consistency when an external failure occurs.
In typical cases a fetch costs one I/O cycle to consult the key file, and if the key is present, one I/O cycle to read the value.
Usage
Callers must define these parameters when creating a database:
KeySize: The size of a key in bytes.BlockSize: The physical size of a key file record.
The ideal block size matches the sector size or block size of the underlying physical media that holds the key file. Functions are provided to return a best estimate of this value for a particular device, but a default of 4096 should work for typical installations. The implementation tries to fit as many entries as possible in a key file record, to maximize the amount of useful work performed per I/O.
LoadFactor: The desired fraction of bucket occupancy
LoadFactor is chosen to make bucket overflows unlikely without
sacrificing bucket occupancy. A value of 0.50 seems to work well with
a good hash function.
Callers must also provide these parameters when a database is opened:
Appnum: An application-defined integer constant which can be retrieved later from the database [TODO].AllocSize: A significant multiple of the average data size.
Memory is recycled to improve performance, so NuDB needs AllocSize as a
hint about the average size of the data being inserted. For an average data size
of 1KB (one kilobyte), AllocSize of sixteen megabytes (16MB) is sufficient. If
the AllocSize is too low, the memory recycler will not make efficient use of
allocated blocks.
Two operations are defined: fetch, and insert.
fetch
The fetch operation retrieves a variable length value given the
key. The caller supplies a factory used to provide a buffer for storing
the value. This interface allows custom memory allocation strategies.
insert
insert adds a key/value pair to the store. Value data must contain at least
one byte. Duplicate keys are disallowed. Insertions are serialized, which means
[TODO].
Implementation
All insertions are buffered in memory, with inserted values becoming immediately discoverable in subsequent or concurrent calls to fetch. Periodically, buffered data is safely committed to disk files using a separate dedicated thread associated with the database. This commit process takes place at least once per second, or more often during a detected surge in insertion activity. In the commit process the key/value pairs receive the following treatment:
An insertion is performed by appending a value record to the data file. The value record has some header information including the size of the data and a copy of the key; the data file is iteratable without the key file. The value data follows the header. The data file is append-only and immutable: once written, bytes are never changed.
Initially the hash table in the key file consists of a single bucket. After the load factor is exceeded from insertions, the hash table grows in size by one bucket by doing a "split". The split operation is the linear hashing algorithm as described by Litwin and Larson.
When a bucket is split, each key is rehashed, and either remains in the original bucket or gets moved to the a bucket appended to the end of the key file.
An insertion on a full bucket first triggers the "spill" algorithm.
First, a spill record is appended to the data file, containing header information followed by the entire bucket record. Then the bucket's size is set to zero and the offset of the spill record is stored in the bucket. At this point the insertion may proceed normally, since the bucket is empty. Spilled buckets in the data file are always full.
Because every bucket holds the offset of the next spill record in the data file, the buckets form a linked list. In practice, careful selection of capacity and load factor will keep the percentage of buckets with one spill record to a minimum, with no bucket requiring two spill records.
The implementation of fetch is straightforward: first the bucket in the key file is checked, then each spill record in the linked list of spill records is checked, until the key is found or there are no more records. As almost all buckets have no spill records, the average fetch requires one I/O (not including reading the value).
One complication in the scheme is when a split occurs on a bucket that has one or more spill records. In this case, both the bucket being split and the new bucket may overflow. This is handled by performing the spill algorithm for each overflow that occurs. The new buckets may have one or more spill records each, depending on the number of keys that were originally present.
Because the data file is immutable, a bucket's original spill records are no longer referenced after the bucket is split. These blocks of data in the data file are unrecoverable wasted space. Correctly configured databases can have a typical waste factor of 1%, which is acceptable. These unused bytes can be removed by visiting each value in the value file using an off-line process and inserting it into a new database, then delete the old database and use the new one instead.
Recovery
To provide atomicity and consistency, a log file associated with the database stores information used to roll back partial commits.
Iteration
Each record in the data file is prefixed with a header identifying whether it is a value record or a spill record, along with the size of the record in bytes and a copy of the key if it's a value record, so values can be iterated by incrementing a byte counter. A key file can be regenerated from just the data file by iterating the values and performing the key insertion algorithm.
Concurrency
Locks are never held during disk reads and writes. Fetches are fully concurrent, while inserts are serialized. Inserts fail on duplicate keys, and are atomic: they either succeed immediately or fail. After an insert, the key is immediately visible to subsequent fetches.
Formats
All integer values are stored as big endian. The uint48_t format consists of 6 bytes.
Key File
The Key File contains the Header followed by one or more fixed-length Bucket Records.
Header (104 bytes)
char[8] Type The characters "nudb.key"
uint16 Version Holds the version number
uint64 UID Unique ID generated on creation
uint64 Appnum Application defined constant
uint16 KeySize Key size in bytes
uint64 Salt A random seed
uint64 Pepper The salt hashed
uint16 BlockSize Size of a file block in bytes
uint16 LoadFactor Target fraction in 65536ths
uint8[56] Reserved Zeroes
uint8[] Reserved Zero-pad to block size
Type identifies the file as belonging to nudb. UID is
generated randomly when the database is created, and this value
is stored in the data and log files as well - it's used
to determine if files belong to the same database. Salt is
generated when the database is created and helps prevent
complexity attacks; it is prepended to the key material
when computing a hash, or used to initialize the state of
the hash function. Appnum is an application defined constant
set when the database is created. It can be used for anything,
for example to distinguish between different data formats.
Pepper is computed by hashing Salt using a hash function
seeded with the salt. This is used to fingerprint the hash
function used. If a database is opened and the fingerprint
does not match the hash calculation performed using the template
argument provided when constructing the store, an exception
is thrown.
The header for the key file contains the File Header followed by the information above. The Capacity is the number of keys per bucket, and defines the size of a bucket record. The load factor is the target fraction of bucket occupancy.
None of the information in the key file header or the data file header may be changed after the database is created, including the Appnum.
Bucket Record (fixed-length)
uint16 Count Number of keys in this bucket
uint48 Spill Offset of the next spill record or 0
BucketEntry[] Entries The bucket entries
Bucket Entry
uint48 Offset Offset in data file of the data
uint48 Size The size of the value in bytes
uint48 Hash The hash of the key
Data File
The Data File contains the Header followed by zero or more variable-length Value Records and Spill Records.
Header (92 bytes)
char[8] Type The characters "nudb.dat"
uint16 Version Holds the version number
uint64 UID Unique ID generated on creation
uint64 Appnum Application defined constant
uint16 KeySize Key size in bytes
uint8[64] (reserved) Zeroes
UID contains the same value as the salt in the corresponding key file. This is placed in the data file so that key and value files belonging to the same database can be identified.
Data Record (variable-length)
uint48 Size Size of the value in bytes
uint8[KeySize] Key The key.
uint8[Size] Data The value data.
Spill Record (fixed-length)
uint48 Zero All zero, identifies a spill record
uint16 Size Bytes in spill bucket (for skipping)
Bucket SpillBucket Bucket Record
Log File
The Log file contains the Header followed by zero or more fixed size log records. Each log record contains a snapshot of a bucket. When a database is not closed cleanly, the recovery process applies the log records to the key file, overwriting data that may be only partially updated with known good information. After the log records are applied, the data and key files are truncated to the last known good size.
Header (62 bytes)
char[8] Type The characters "nudb.log"
uint16 Version Holds the version number
uint64 UID Unique ID generated on creation
uint64 Appnum Application defined constant
uint16 KeySize Key size in bytes
uint64 Salt A random seed.
uint64 Pepper The salt hashed
uint16 BlockSize Size of a file block in bytes
uint64 KeyFileSize Size of key file.
uint64 DataFileSize Size of data file.
Log Record
uint64_t Index Bucket index (0-based)
Bucket Bucket Compact Bucket record
Compact buckets include only Size entries. These are primarily used to minimize the volume of writes to the log file.
License
Distributed under the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
Contact
Please report issues or questions here: https://github.com/vinniefalco/NuDB/issues