Legacy workarounds for Visual Studio non thread-safe initialization
of function local objects with static storage duration are removed:
* Remove LeakChecked
* Remove StaticObject
* Remove SharedSingleton
This adds codecs for snappy and lz4, and a new nodeobject codec. The
nodeobject codec provides a highly efficient custom compression scheme
for inner nodes, which make up the majority of nodestore databases.
Non inner node objects are compressed using lz4.
The NuDB backend is modified to use the nodeobject codec. This change
is not backward compatible - older NuDB databases cannot be opened or
imported.
This introduces changes in nudb to improve speed, reduce database size,
and enhance correctness. The most significant change is to store hashes
rather than entire keys in the key file. The output of the hash function
is reduced to 48 bits, and stored directly in buckets.
The API is also modified to introduce a Codec parameter allowing for
compression and decompression to be supported in the implementation
itself rather than callers.
THe data file no longer contains a salt, as the salt is applicable
only to the key and log files. This allows a data file to have multiple
key files with different salt values. To distinguish physical files
belonging to the same logical database, a new field UID is introduced.
The UID is a 64-bit random value generated once on creation and stored
in all three files.
Buckets are zero filled to the end of each block, this is a security
measure to prevent unintended contents of memory getting stored to
disk. NuDB offers the varint integer type, this is identical to
the varint described by Google.
* Add varint
* Add Codec template argument
* Add "api" convenience traits
* Store hash in buckets
* istream can throw short read errors
* Support std::uint8_t format in streams
* Make file classes part of the public interface
* Remove buffers pessimization, replace with buffer
* Consolidate creation utility functions to the same header
* Zero fill unused areas of buckets on disk
* More coverage and improvements to the recover test
* Fix file read/write to loop until all bytes processed
* Add verify_fast, faster verify for large databases
The database version number is incremented to 2; older databases can
no longer be opened and should be deleted.
This reverts the change that makes RocksDBQuick the default settings for
node_db "type=rocksdb". The quick settings can be obtained by setting
"type=rocksdbquick".
RocksDBQuick settings are implicated in memory over-utilization problems
seen recently.
* Remove unused members
* SerialIter holds only a pointer and offset now
* Use free functions for some Serializer members
* Use SerialIter in some places instead of Serializer
The LevelDB and HyperLevelDB are removed from the backend choices. Neither
were recommended for production environments. As RocksDB is not available
on Windows platforms yet, the recommended backend choice for Windows is NuDB.
Add metrics to record the number of RPC requests received. Record the number of
node store fetches performed per request. Additionally record the byte size of
each request response and measure the response time of each request in
milliseconds.
A new class, ScopedMetrics, uses the Boost Thead Local Storage mechanism to
efficiently record NodeStore metrics within the same thread.
The NuDB database backend is a high performance key/value store presented
as an alternative to RocksDB on Mac and Linux deployments, and the preferred
backend option for Windows deployments. The LevelDB backend is deprecated for
all platforms.
This includes these changes:
* Add Backend::verify API for doing consistency checks
* Add Database::close so caller can catch exceptions
* Improved Timing test for NodeStore creates a simulated workload
* Add Backend::verify API for doing consistency checks
* Add Database::close so caller can catch exceptions
* Improved Timing test for NodeStore creates a simulated workload
This removes the old default configuration for the "rocksdb" backend and
replaces it with the configuration that was formerly available using
the experimental backend "rocksdbquick".
The new configuration setting improves the performance of the key/value
database by changing the compaction style and tuning the size parameters for
the typical rippled workload. Testing shows a decrease in I/O spikes for both
reading and writing.
An alternative to the unity build, the classic build compiles each
translation unit individually. This adds more modules to the classic build:
* Remove unity header app.h
* Add missing includes as needed
* Remove obsolete NodeStore backend code
* Add app/, core/, crypto/, json/, net/, overlay/, peerfinder/ to classic build
Source files are split to place all unit test code into translation
units ending in .test.cpp with no other business logic in the same file,
and in directories named "test".
A new target is added to the SConstruct, invoked by:
scons count
This prints the total number of source code lines occupied by unit tests,
in rippled specific code and excluding library subtrees.
These identifiers were part of a failed set of classes to replace
the functionality combined into RippleAddress. They are not used
and therefore can be removed.
* Remove RippleAccountPrivateKey
* Remove RippleAccountPublicKey
* Remove RippleAccountID
* Remove RipplePrivateKey
* Remove RipplePublicKeyHash
* Remove RippleLedgerHash
* Remove unused withCheck argument
* Remove CryptoIdentifier
* Remove IdentifierStorage
* Remove IdentifierType
* Remove SimpleIdentifier
* Add missing include
Makes rippled configurable to support deletion of all data in its key-value
store (nodestore) and ledger and transaction SQLite databases based on
validated ledger sequence numbers. All records from a specified ledger
and forward shall remain available in the key-value store and SQLite, and
all data prior to that specific ledger may be deleted.
Additionally, the administrator may require that an RPC command be
executed to enable deletion. This is to align data deletion with local
policy.
* New src/ripple/crypto and src/ripple/protocol directories
* Merged src/ripple/common into src/ripple/basics
* Move resource/api files up a level
* Add headers for "include what you use"
* Normalized include guards
* Renamed to JsonFields.h
* Remove obsolete files
* Remove net.h unity header
* Remove resource.h unity header
* Removed some deprecated unity includes
* Use more succinct while loops on NodeFactory.
* Better formatting of multiple test results.
* Updated benchmarks.
* Use simpler and faster RNG to generate test data.
This new factory is intended for benchmarking against the existing RocksDBFactory and has the following differences.
* Does not use BatchWriter
* Disables WAL for writes to memtable
* Uses a hash index in blocks
* Uses RocksDB OptimizeFor… functions
See Benchmarks.md for further discussion of some of the issues raised by investigation of RocksDB performance.
The timing test is changed to overcome possible file buffer cache effects by creating different read access patterns. The unittest-arg command line arguments allow running the benchmarks against any of the available backends and altering the parameters passed in the same format as rippled.cfg. The num_objects parameter permits variation of the number of key/values inserted. The data is random but matches reasonably well the values that rippled might generate.
This seemed to improve the performance of the copy, although there did seem to be some byte by byte copying still present. Further investigation recommended.
* Use static_assert where appropriate
* Use std::min and std::max where appropriate
* Simplify RippleD error reporting
* Remove use of beast::RandomAccessFile
Empirical evidence shows a database access pattern with few hits
and many misses (objects that don't exist). This changes the timing
tests so they more accurately reflect rippled's actual usage:
* Add read missing keys test
* Increase numObjectsToTest to 1,000,000
* Alter PredictableObjectFactory to seed RNG once only
* Make NodeStoreTiming a manual test
