engine

package
v1.0.5 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Aug 17, 2017 License: Apache-2.0 Imports: 33 Imported by: 0

Documentation

Overview

Package engine provides low-level storage. It interacts with storage backends (e.g. LevelDB, RocksDB, etc.) via the Engine interface. At one level higher, MVCC provides multi-version concurrency control capability on top of an Engine instance.

The Engine interface provides an API for key-value stores. InMem implements an in-memory engine using a sorted map. RocksDB implements an engine for data stored to local disk using RocksDB, a variant of LevelDB.

MVCC provides a multi-version concurrency control system on top of an engine. MVCC is the basis for Cockroach's support for distributed transactions. It is intended for direct use from storage.Range objects.

Notes on MVCC architecture

Each MVCC value contains a metadata key/value pair and one or more version key/value pairs. The MVCC metadata key is the actual key for the value, using the util/encoding.EncodeBytes scheme. The MVCC metadata value is of type MVCCMetadata and contains the most recent version timestamp and an optional roachpb.Transaction message. If set, the most recent version of the MVCC value is a transactional "intent". It also contains some information on the size of the most recent version's key and value for efficient stat counter computations. Note that it is not necessary to explicitly store the MVCC metadata as its contents can be reconstructed from the most recent versioned value as long as an intent is not present. The implementation takes advantage of this and deletes the MVCC metadata when possible.

Each MVCC version key/value pair has a key which is also binary-encoded, but is suffixed with a decreasing, big-endian encoding of the timestamp (eight bytes for the nanosecond wall time, followed by four bytes for the logical time except for meta key value pairs, for which the timestamp is implicit). The MVCC version value is a message of type roachpb.Value. A deletion is indicated by an empty value. Note that an empty roachpb.Value will encode to a non-empty byte slice. The decreasing encoding on the timestamp sorts the most recent version directly after the metadata key, which is treated specially by the RocksDB comparator (by making the zero timestamp sort first). This increases the likelihood that an Engine.Get() of the MVCC metadata will get the same block containing the most recent version, even if there are many versions. We rely on getting the MVCC metadata key/value and then using it to directly get the MVCC version using the metadata's most recent version timestamp. This avoids using an expensive merge iterator to scan the most recent version. It also allows us to leverage RocksDB's bloom filters.

The binary encoding used on the MVCC keys allows arbitrary keys to be stored in the map (no restrictions on intermediate nil-bytes, for example), while still sorting lexicographically and guaranteeing that all timestamp-suffixed MVCC version keys sort consecutively with the metadata key. We use an escape-based encoding which transforms all nul ("\x00") characters in the key and is terminated with the sequence "\x00\x01", which is guaranteed to not occur elsewhere in the encoded value. See util/encoding/encoding.go for more details.

We considered inlining the most recent MVCC version in the MVCCMetadata. This would reduce the storage overhead of storing the same key twice (which is small due to block compression), and the runtime overhead of two separate DB lookups. On the other hand, all writes that create a new version of an existing key would incur a double write as the previous value is moved out of the MVCCMetadata into its versioned key. Preliminary benchmarks have not shown enough performance improvement to justify this change, although we may revisit this decision if it turns out that multiple versions of the same key are rare in practice.

However, we do allow inlining in order to use the MVCC interface to store non-versioned values. It turns out that not everything which Cockroach needs to store would be efficient or possible using MVCC. Examples include transaction records, response cache entries, stats counters, time series data, and system-local config values. However, supporting a mix of encodings is problematic in terms of resulting complexity. So Cockroach treats an MVCC timestamp of zero to mean an inlined, non-versioned value. These values are replaced if they exist on a Put operation and are cleared from the engine on a delete. Importantly, zero-timestamped MVCC values may be merged, as is necessary for stats counters and time series data.

Index

Constants

View Source
const (
	BatchTypeDeletion BatchType = 0x0
	BatchTypeValue              = 0x1
	BatchTypeMerge              = 0x2
)

These constants come from rocksdb/db/dbformat.h.

View Source
const (

	// DefaultMaxOpenFiles is the default value for rocksDB's max_open_files
	// option.
	DefaultMaxOpenFiles = -1
	// RecommendedMaxOpenFiles is the recommended value for rocksDB's
	// max_open_files option. If more file descriptors are available than the
	// recommended number, than the default value is used.
	RecommendedMaxOpenFiles = 10000
	// MinimumMaxOpenFiles is The minimum value that rocksDB's max_open_files
	// option can be set to. While this should be set as high as possible, the
	// minimum total for a single store node must be under 2048 for Windows
	// compatibility. See:
	// https://wpdev.uservoice.com/forums/266908-command-prompt-console-bash-on-ubuntu-on-windo/suggestions/17310124-add-ability-to-change-max-number-of-open-files-for
	MinimumMaxOpenFiles = 1700
)

Variables

View Source
var (
	// MVCCKeyMax is a maximum mvcc-encoded key value which sorts after
	// all other keys.`
	MVCCKeyMax = MakeMVCCMetadataKey(roachpb.KeyMax)
	// NilKey is the nil MVCCKey.
	NilKey = MVCCKey{}
)

Functions

func AccountForSelf

func AccountForSelf(ms *enginepb.MVCCStats, rangeID roachpb.RangeID) error

AccountForSelf adjusts ms to account for the predicted impact it will have on the values that it records when the structure is initially stored. Specifically, MVCCStats is stored on the RangeStats key, which means that its creation will have an impact on system-local data size and key count.

func IsIntentOf

func IsIntentOf(meta enginepb.MVCCMetadata, txn *roachpb.Transaction) bool

IsIntentOf returns true if the meta record is an intent of the supplied transaction.

func IsValidSplitKey

func IsValidSplitKey(key roachpb.Key) bool

IsValidSplitKey returns whether the key is a valid split key. Certain key ranges cannot be split (the meta1 span and the system DB span); split keys chosen within any of these ranges are considered invalid. And a split key equal to Meta2KeyMax (\x03\xff\xff) is considered invalid.

func MVCCBlindConditionalPut

func MVCCBlindConditionalPut(
	ctx context.Context,
	engine Writer,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
	expVal *roachpb.Value,
	txn *roachpb.Transaction,
) error

MVCCBlindConditionalPut is a fast-path of MVCCConditionalPut. See the MVCCConditionalPut comments for details of the semantics. MVCCBlindConditionalPut skips retrieving the existing metadata for the key requiring the caller to guarantee no versions for the key currently exist.

func MVCCBlindInitPut

func MVCCBlindInitPut(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
	txn *roachpb.Transaction,
) error

MVCCBlindInitPut is a fast-path of MVCCInitPut. See the MVCCInitPut comments for details of the semantics. MVCCBlindInitPut skips retrieving the existing metadata for the key requiring the caller to gauarntee no version for the key currently exist.

func MVCCBlindPut

func MVCCBlindPut(
	ctx context.Context,
	engine Writer,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
	txn *roachpb.Transaction,
) error

MVCCBlindPut is a fast-path of MVCCPut. See the MVCCPut comments for details of the semantics. MVCCBlindPut skips retrieving the existing metadata for the key requiring the caller to guarantee no versions for the key currently exist in order for stats to be updated properly. If a previous version of the key does exist it is up to the caller to properly account for their existence in updating the stats.

func MVCCConditionalPut

func MVCCConditionalPut(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
	expVal *roachpb.Value,
	txn *roachpb.Transaction,
) error

MVCCConditionalPut sets the value for a specified key only if the expected value matches. If not, the return a ConditionFailedError containing the actual value.

The condition check reads a value from the key using the same operational timestamp as we use to write a value.

func MVCCDelete

func MVCCDelete(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	txn *roachpb.Transaction,
) error

MVCCDelete marks the key deleted so that it will not be returned in future get responses.

func MVCCDeleteRange

func MVCCDeleteRange(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key,
	endKey roachpb.Key,
	max int64,
	timestamp hlc.Timestamp,
	txn *roachpb.Transaction,
	returnKeys bool,
) ([]roachpb.Key, *roachpb.Span, int64, error)

MVCCDeleteRange deletes the range of key/value pairs specified by start and end keys. It returns the range of keys deleted when returnedKeys is set, the next span to resume from, and the number of keys deleted.

func MVCCFindSplitKey

func MVCCFindSplitKey(
	ctx context.Context,
	engine Reader,
	rangeID roachpb.RangeID,
	key,
	endKey roachpb.RKey,
	targetSize int64,
) (roachpb.Key, error)

MVCCFindSplitKey suggests a split key from the given user-space key range that aims to roughly cut into half the total number of bytes used (in raw key and value byte strings) in both subranges. Specify a snapshot engine to safely invoke this method in a goroutine.

The split key will never be chosen from the key ranges listed in illegalSplitKeySpans.

debugFn, if not nil, is used to print informational log messages about the key finding process.

func MVCCGarbageCollect

func MVCCGarbageCollect(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	keys []roachpb.GCRequest_GCKey,
	timestamp hlc.Timestamp,
	maxClears int64,
) error

MVCCGarbageCollect creates an iterator on the engine. In parallel it iterates through the keys listed for garbage collection by the keys slice. The engine iterator is seeked in turn to each listed key, clearing all values with timestamps <= to expiration. The timestamp parameter is used to compute the intent age on GC. Garbage collection stops after clearing maxClears values (to limit the size of the WriteBatch produced).

func MVCCGet

func MVCCGet(
	ctx context.Context,
	engine Reader,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	consistent bool,
	txn *roachpb.Transaction,
) (*roachpb.Value, []roachpb.Intent, error)

MVCCGet returns the value for the key specified in the request, while satisfying the given timestamp condition. The key may contain arbitrary bytes. If no value for the key exists, or it has been deleted, returns nil for value.

The values of multiple versions for the given key should be organized as follows: ... keyA : MVCCMetadata of keyA keyA_Timestamp_n : value of version_n keyA_Timestamp_n-1 : value of version_n-1 ... keyA_Timestamp_0 : value of version_0 keyB : MVCCMetadata of keyB ...

The consistent parameter indicates that intents should cause WriteIntentErrors. If set to false, a possible intent on the key will be ignored for reading the value (but returned via the roachpb.Intent slice); the previous value (if any) is read instead.

func MVCCGetAsTxn

func MVCCGetAsTxn(
	ctx context.Context,
	engine Reader,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	txnMeta enginepb.TxnMeta,
) (*roachpb.Value, []roachpb.Intent, error)

MVCCGetAsTxn constructs a temporary Transaction from the given txn metadata and calls MVCCGet as that transaction. This method is required only for reading intents of a transaction when only its metadata is known and should rarely be used. The read is carried out without the chance of uncertainty restarts.

func MVCCGetProto

func MVCCGetProto(
	ctx context.Context,
	engine Reader,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	consistent bool,
	txn *roachpb.Transaction,
	msg proto.Message,
) (bool, error)

MVCCGetProto fetches the value at the specified key and unmarshals it into msg if msg is non-nil. Returns true on success or false if the key was not found. The semantics of consistent are the same as in MVCCGet.

func MVCCGetRangeStats

func MVCCGetRangeStats(
	ctx context.Context, engine Reader, rangeID roachpb.RangeID,
) (enginepb.MVCCStats, error)

MVCCGetRangeStats reads stat counters for the specified range and sets the values in the enginepb.MVCCStats struct.

func MVCCIncrement

func MVCCIncrement(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	txn *roachpb.Transaction,
	inc int64,
) (int64, error)

MVCCIncrement fetches the value for key, and assuming the value is an "integer" type, increments it by inc and stores the new value. The newly incremented value is returned.

An initial value is read from the key using the same operational timestamp as we use to write a value.

func MVCCInitPut

func MVCCInitPut(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
	txn *roachpb.Transaction,
) error

MVCCInitPut sets the value for a specified key if the key doesn't exist. It returns an error when the write fails or if the key exists with an existing value that is different from the supplied value.

func MVCCIterate

func MVCCIterate(
	ctx context.Context,
	engine Reader,
	startKey,
	endKey roachpb.Key,
	timestamp hlc.Timestamp,
	consistent bool,
	txn *roachpb.Transaction,
	reverse bool,
	f func(roachpb.KeyValue) (bool, error),
) ([]roachpb.Intent, error)

MVCCIterate iterates over the key range [start,end). At each step of the iteration, f() is invoked with the current key/value pair. If f returns true (done) or an error, the iteration stops and the error is propagated. If the reverse is flag set the iterator will be moved in reverse order.

func MVCCMerge

func MVCCMerge(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
) error

MVCCMerge implements a merge operation. Merge adds integer values, concatenates undifferentiated byte slice values, and efficiently combines time series observations if the roachpb.Value tag value indicates the value byte slice is of type TIMESERIES.

func MVCCPut

func MVCCPut(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	value roachpb.Value,
	txn *roachpb.Transaction,
) error

MVCCPut sets the value for a specified key. It will save the value with different versions according to its timestamp and update the key metadata. The timestamp must be passed as a parameter; using the Timestamp field on the value results in an error.

If the timestamp is specified as hlc.Timestamp{}, the value is inlined instead of being written as a timestamp-versioned value. A zero timestamp write to a key precludes a subsequent write using a non-zero timestamp and vice versa. Inlined values require only a single row and never accumulate more than a single value. Successive zero timestamp writes to a key replace the value and deletes clear the value. In addition, zero timestamp values may be merged.

func MVCCPutProto

func MVCCPutProto(
	ctx context.Context,
	engine ReadWriter,
	ms *enginepb.MVCCStats,
	key roachpb.Key,
	timestamp hlc.Timestamp,
	txn *roachpb.Transaction,
	msg proto.Message,
) error

MVCCPutProto sets the given key to the protobuf-serialized byte string of msg and the provided timestamp.

func MVCCResolveWriteIntent

func MVCCResolveWriteIntent(
	ctx context.Context, engine ReadWriter, ms *enginepb.MVCCStats, intent roachpb.Intent,
) error

MVCCResolveWriteIntent either commits or aborts (rolls back) an extant write intent for a given txn according to commit parameter. ResolveWriteIntent will skip write intents of other txns.

Transaction epochs deserve a bit of explanation. The epoch for a transaction is incremented on transaction retries. A transaction retry is different from an abort. Retries can occur in SSI transactions when the commit timestamp is not equal to the proposed transaction timestamp. On a retry, the epoch is incremented instead of creating an entirely new transaction. This allows the intents that were written on previous runs to serve as locks which prevent concurrent reads from further incrementing the timestamp cache, making further transaction retries less likely.

Because successive retries of a transaction may end up writing to different keys, the epochs serve to classify which intents get committed in the event the transaction succeeds (all those with epoch matching the commit epoch), and which intents get aborted, even if the transaction succeeds.

TODO(tschottdorf): encountered a bug in which a Txn committed with its original timestamp after laying down intents at higher timestamps. Doesn't look like this code here caught that. Shouldn't resolve intents when they're not at the timestamp the Txn mandates them to be.

func MVCCResolveWriteIntentRange

func MVCCResolveWriteIntentRange(
	ctx context.Context, engine ReadWriter, ms *enginepb.MVCCStats, intent roachpb.Intent, max int64,
) (int64, error)

MVCCResolveWriteIntentRange commits or aborts (rolls back) the range of write intents specified by start and end keys for a given txn. ResolveWriteIntentRange will skip write intents of other txns.

func MVCCResolveWriteIntentRangeUsingIter

func MVCCResolveWriteIntentRangeUsingIter(
	ctx context.Context,
	engine ReadWriter,
	iterAndBuf IterAndBuf,
	ms *enginepb.MVCCStats,
	intent roachpb.Intent,
	max int64,
) (int64, error)

MVCCResolveWriteIntentRangeUsingIter commits or aborts (rolls back) the range of write intents specified by start and end keys for a given txn. ResolveWriteIntentRange will skip write intents of other txns.

func MVCCResolveWriteIntentUsingIter

func MVCCResolveWriteIntentUsingIter(
	ctx context.Context,
	engine ReadWriter,
	iterAndBuf IterAndBuf,
	ms *enginepb.MVCCStats,
	intent roachpb.Intent,
) error

MVCCResolveWriteIntentUsingIter is a variant of MVCCResolveWriteIntent that uses iterator and buffer passed as parameters (e.g. when used in a loop).

func MVCCReverseScan

func MVCCReverseScan(
	ctx context.Context,
	engine Reader,
	key,
	endKey roachpb.Key,
	max int64,
	timestamp hlc.Timestamp,
	consistent bool,
	txn *roachpb.Transaction,
) ([]roachpb.KeyValue, *roachpb.Span, []roachpb.Intent, error)

MVCCReverseScan scans the key range [start,end) key up to some maximum number of results in descending order. If it hits max, it returns a span to be used in the next call to this function.

func MVCCScan

func MVCCScan(
	ctx context.Context,
	engine Reader,
	key,
	endKey roachpb.Key,
	max int64,
	timestamp hlc.Timestamp,
	consistent bool,
	txn *roachpb.Transaction,
) ([]roachpb.KeyValue, *roachpb.Span, []roachpb.Intent, error)

MVCCScan scans the key range [start,end) key up to some maximum number of results in ascending order. If it hits max, it returns a span to be used in the next call to this function.

func MVCCSetRangeStats

func MVCCSetRangeStats(
	ctx context.Context, engine ReadWriter, rangeID roachpb.RangeID, ms *enginepb.MVCCStats,
) error

MVCCSetRangeStats sets stat counters for specified range.

func MakeValue

func MakeValue(meta enginepb.MVCCMetadata) roachpb.Value

MakeValue returns the inline value.

func MergeInternalTimeSeriesData

func MergeInternalTimeSeriesData(
	sources ...roachpb.InternalTimeSeriesData,
) (roachpb.InternalTimeSeriesData, error)

MergeInternalTimeSeriesData exports the engine's C++ merge logic for InternalTimeSeriesData to higher level packages. This is intended primarily for consumption by high level testing of time series functionality.

func PutProto

func PutProto(engine Writer, key MVCCKey, msg proto.Message) (keyBytes, valBytes int64, err error)

PutProto sets the given key to the protobuf-serialized byte string of msg and the provided timestamp. Returns the length in bytes of key and the value.

func RunLDB

func RunLDB(args []string)

RunLDB runs RocksDB's ldb command-line tool. The passed command-line arguments should not include argv[0].

Types

type Batch

type Batch interface {
	ReadWriter
	// Commit atomically applies any batched updates to the underlying
	// engine. This is a noop unless the engine was created via NewBatch(). If
	// sync is true, the batch is synchronously committed to disk.
	Commit(sync bool) error
	// Distinct returns a view of the existing batch which only sees writes that
	// were performed before the Distinct batch was created. That is, the
	// returned batch will not read its own writes, but it will read writes to
	// the parent batch performed before the call to Distinct(). The returned
	// batch needs to be closed before using the parent batch again. This is used
	// as an optimization to avoid flushing mutations buffered by the batch in
	// situations where we know all of the batched operations are for distinct
	// keys.
	Distinct() ReadWriter
	// Repr returns the underlying representation of the batch and can be used to
	// reconstitute the batch on a remote node using Writer.ApplyBatchRepr().
	Repr() []byte
}

Batch is the interface for batch specific operations.

type BatchType

type BatchType byte

BatchType represents the type of an entry in an encoded RocksDB batch.

type Engine

type Engine interface {
	ReadWriter
	// Attrs returns the engine/store attributes.
	Attrs() roachpb.Attributes
	// Capacity returns capacity details for the engine's available storage.
	Capacity() (roachpb.StoreCapacity, error)
	// Flush causes the engine to write all in-memory data to disk
	// immediately.
	Flush() error
	// GetStats retrieves stats from the engine.
	GetStats() (*Stats, error)
	// GetTempDir returns a path under which tempdirs or tempfiles can be created.
	GetTempDir() string
	// NewBatch returns a new instance of a batched engine which wraps
	// this engine. Batched engines accumulate all mutations and apply
	// them atomically on a call to Commit().
	NewBatch() Batch
	// NewWriteOnlyBatch returns a new instance of a batched engine which wraps
	// this engine. A write-only batch accumulates all mutations and applies them
	// atomically on a call to Commit(). Read operations return an error.
	//
	// TODO(peter): This should return a WriteBatch interface, but there are mild
	// complications in both defining that interface and implementing it. In
	// particular, Batch.Close would no longer come from Reader and we'd need to
	// refactor a bunch of code in rocksDBBatch.
	NewWriteOnlyBatch() Batch
	// NewSnapshot returns a new instance of a read-only snapshot
	// engine. Snapshots are instantaneous and, as long as they're
	// released relatively quickly, inexpensive. Snapshots are released
	// by invoking Close(). Note that snapshots must not be used after the
	// original engine has been stopped.
	NewSnapshot() Reader
	// SetTempDir overrides the tempdir path returned by GetTempDir.
	SetTempDir(dir string) error
}

Engine is the interface that wraps the core operations of a key/value store.

type GarbageCollector

type GarbageCollector struct {
	Threshold hlc.Timestamp
	// contains filtered or unexported fields
}

GarbageCollector GCs MVCC key/values using a zone-specific GC policy allows either the union or intersection of maximum # of versions and maximum age.

func MakeGarbageCollector

func MakeGarbageCollector(now hlc.Timestamp, policy config.GCPolicy) GarbageCollector

MakeGarbageCollector allocates and returns a new GC, with expiration computed based on current time and policy.TTLSeconds.

func (GarbageCollector) Filter

func (gc GarbageCollector) Filter(keys []MVCCKey, values [][]byte) hlc.Timestamp

Filter makes decisions about garbage collection based on the garbage collection policy for batches of values for the same key. Returns the timestamp including, and after which, all values should be garbage collected. If no values should be GC'd, returns the zero timestamp. keys must be in descending time order. Values deleted at or before the returned timestamp can be deleted without invalidating any reads in the time interval (gc.expiration, \infinity).

The GC keeps all values (including deletes) above the expiration time, plus the first value before or at the expiration time. This allows reads to be guaranteed as described above. However if this were the only rule, then if the most recent write was a delete, it would never be removed. Thus, when a deleted value is the most recent before expiration, it can be deleted. This would still allow for the tombstone bugs in #6227, so in the future we will add checks that disallow writes before the last GC expiration time.

type InMem

type InMem struct {
	*RocksDB
}

InMem wraps RocksDB and configures it for in-memory only storage.

func NewInMem

func NewInMem(attrs roachpb.Attributes, cacheSize int64) InMem

NewInMem allocates and returns a new, opened InMem engine. The caller must call the engine's Close method when the engine is no longer needed.

type IterAndBuf

type IterAndBuf struct {
	// contains filtered or unexported fields
}

IterAndBuf used to pass iterators and buffers between MVCC* calls, allowing reuse without the callers needing to know the particulars.

func GetIterAndBuf

func GetIterAndBuf(engine Reader) IterAndBuf

GetIterAndBuf returns a IterAndBuf for passing into various MVCC* methods.

func (IterAndBuf) Cleanup

func (b IterAndBuf) Cleanup()

Cleanup must be called to release the resources when done.

type Iterator

type Iterator interface {
	// Close frees up resources held by the iterator.
	Close()
	// Seek advances the iterator to the first key in the engine which
	// is >= the provided key.
	Seek(key MVCCKey)
	// SeekReverse advances the iterator to the first key in the engine which
	// is <= the provided key.
	SeekReverse(key MVCCKey)
	// Valid must be called after any call to Seek(), Next(), Prev(), or
	// similar methods. It returns (true, nil) if the iterator points to
	// a valid key (it is undefined to call Key(), Value(), or similar
	// methods unless Valid() has returned (true, nil)). It returns
	// (false, nil) if the iterator has moved past the end of the valid
	// range, or (false, err) if an error has occurred. Valid() will
	// never return true with a non-nil error.
	Valid() (bool, error)
	// Next advances the iterator to the next key/value in the
	// iteration. After this call, Valid() will be true if the
	// iterator was not positioned at the last key.
	Next()
	// Prev moves the iterator backward to the previous key/value
	// in the iteration. After this call, Valid() will be true if the
	// iterator was not positioned at the first key.
	Prev()
	// NextKey advances the iterator to the next MVCC key. This operation is
	// distinct from Next which advances to the next version of the current key
	// or the next key if the iterator is currently located at the last version
	// for a key.
	NextKey()
	// PrevKey moves the iterator backward to the previous MVCC key. This
	// operation is distinct from Prev which moves the iterator backward to the
	// prev version of the current key or the prev key if the iterator is
	// currently located at the first version for a key.
	PrevKey()
	// Key returns the current key.
	Key() MVCCKey
	// Value returns the current value as a byte slice.
	Value() []byte
	// ValueProto unmarshals the value the iterator is currently
	// pointing to using a protobuf decoder.
	ValueProto(msg proto.Message) error
	// UnsafeKey returns the same value as Key, but the memory is invalidated on
	// the next call to {Next,Prev,Seek,SeekReverse,Close}.
	UnsafeKey() MVCCKey
	// UnsafeValue returns the same value as Value, but the memory is
	// invalidated on the next call to {Next,Prev,Seek,SeekReverse,Close}.
	UnsafeValue() []byte
	// Less returns true if the key the iterator is currently positioned at is
	// less than the specified key.
	Less(key MVCCKey) bool
	// ComputeStats scans the underlying engine from start to end keys and
	// computes stats counters based on the values. This method is used after a
	// range is split to recompute stats for each subrange. The start key is
	// always adjusted to avoid counting local keys in the event stats are being
	// recomputed for the first range (i.e. the one with start key == KeyMin).
	// The nowNanos arg specifies the wall time in nanoseconds since the
	// epoch and is used to compute the total age of all intents.
	ComputeStats(start, end MVCCKey, nowNanos int64) (enginepb.MVCCStats, error)
}

Iterator is an interface for iterating over key/value pairs in an engine. Iterator implementations are thread safe unless otherwise noted.

type MVCCKey

type MVCCKey struct {
	Key       roachpb.Key
	Timestamp hlc.Timestamp
}

MVCCKey is a versioned key, distinguished from roachpb.Key with the addition of a timestamp.

func AllocIterKeyValue

func AllocIterKeyValue(
	a bufalloc.ByteAllocator, iter Iterator,
) (bufalloc.ByteAllocator, MVCCKey, []byte)

AllocIterKeyValue returns iter.Key() and iter.Value() with the underlying storage allocated from the passed ChunkAllocator.

func DecodeKey

func DecodeKey(encodedKey []byte) (MVCCKey, error)

DecodeKey decodes an engine.MVCCKey from its serialized representation. This decoding must match engine/db.cc:DecodeKey().

func MakeMVCCMetadataKey

func MakeMVCCMetadataKey(key roachpb.Key) MVCCKey

MakeMVCCMetadataKey creates an MVCCKey from a roachpb.Key.

func (MVCCKey) EncodedSize

func (k MVCCKey) EncodedSize() int

EncodedSize returns the size of the MVCCKey when encoded.

func (MVCCKey) Equal

func (k MVCCKey) Equal(l MVCCKey) bool

Equal returns whether two keys are identical.

func (MVCCKey) IsValue

func (k MVCCKey) IsValue() bool

IsValue returns true iff the timestamp is non-zero.

func (MVCCKey) Less

func (k MVCCKey) Less(l MVCCKey) bool

Less compares two keys.

func (MVCCKey) Next

func (k MVCCKey) Next() MVCCKey

Next returns the next key.

func (MVCCKey) String

func (k MVCCKey) String() string

String returns a string-formatted version of the key.

type MVCCKeyValue

type MVCCKeyValue struct {
	Key   MVCCKey
	Value []byte
}

MVCCKeyValue contains the raw bytes of the value for a key.

func Scan

func Scan(engine Reader, start, end MVCCKey, max int64) ([]MVCCKeyValue, error)

Scan returns up to max key/value objects starting from start (inclusive) and ending at end (non-inclusive). Specify max=0 for unbounded scans.

type ReadWriter

type ReadWriter interface {
	Reader
	Writer
}

ReadWriter is the read/write interface to an engine's data.

type Reader

type Reader interface {
	// Close closes the reader, freeing up any outstanding resources. Note that
	// various implementations have slightly different behaviors. In particular,
	// Distinct() batches release their parent batch for future use while
	// Engines, Snapshots and Batches free the associated C++ resources.
	Close()
	// Closed returns true if the reader has been closed or is not usable.
	// Objects backed by this reader (e.g. Iterators) can check this to ensure
	// that they are not using a closed engine. Intended for use within package
	// engine; exported to enable wrappers to exist in other packages.
	Closed() bool
	// Get returns the value for the given key, nil otherwise.
	Get(key MVCCKey) ([]byte, error)
	// GetProto fetches the value at the specified key and unmarshals it
	// using a protobuf decoder. Returns true on success or false if the
	// key was not found. On success, returns the length in bytes of the
	// key and the value.
	GetProto(key MVCCKey, msg proto.Message) (ok bool, keyBytes, valBytes int64, err error)
	// Iterate scans from start to end keys, visiting at most max
	// key/value pairs. On each key value pair, the function f is
	// invoked. If f returns an error or if the scan itself encounters
	// an error, the iteration will stop and return the error.
	// If the first result of f is true, the iteration stops.
	Iterate(start, end MVCCKey, f func(MVCCKeyValue) (bool, error)) error
	// NewIterator returns a new instance of an Iterator over this engine. When
	// prefix is true, Seek will use the user-key prefix of the supplied MVCC key
	// to restrict which sstables are searched, but iteration (using Next) over
	// keys without the same user-key prefix will not work correctly (keys may be
	// skipped). The caller must invoke Iterator.Close() when finished with the
	// iterator to free resources.
	NewIterator(prefix bool) Iterator
}

Reader is the read interface to an engine's data.

type RocksDB

type RocksDB struct {
	// contains filtered or unexported fields
}

RocksDB is a wrapper around a RocksDB database instance.

func NewRocksDB

func NewRocksDB(
	attrs roachpb.Attributes, dir string, cache RocksDBCache, maxSize int64, maxOpenFiles int,
) (*RocksDB, error)

NewRocksDB allocates and returns a new RocksDB object. This creates options and opens the database. If the database doesn't yet exist at the specified directory, one is initialized from scratch. The caller must call the engine's Close method when the engine is no longer needed.

func (*RocksDB) ApplyBatchRepr

func (r *RocksDB) ApplyBatchRepr(repr []byte, sync bool) error

ApplyBatchRepr atomically applies a set of batched updates. Created by calling Repr() on a batch. Using this method is equivalent to constructing and committing a batch whose Repr() equals repr.

func (*RocksDB) Attrs

func (r *RocksDB) Attrs() roachpb.Attributes

Attrs returns the list of attributes describing this engine. This may include a specification of disk type (e.g. hdd, ssd, fio, etc.) and potentially other labels to identify important attributes of the engine.

func (*RocksDB) Capacity

func (r *RocksDB) Capacity() (roachpb.StoreCapacity, error)

Capacity queries the underlying file system for disk capacity information.

func (*RocksDB) Clear

func (r *RocksDB) Clear(key MVCCKey) error

Clear removes the item from the db with the given key.

func (*RocksDB) ClearIterRange

func (r *RocksDB) ClearIterRange(iter Iterator, start, end MVCCKey) error

ClearIterRange removes a set of entries, from start (inclusive) to end (exclusive).

func (*RocksDB) ClearRange

func (r *RocksDB) ClearRange(start, end MVCCKey) error

ClearRange removes a set of entries, from start (inclusive) to end (exclusive).

func (*RocksDB) Close

func (r *RocksDB) Close()

Close closes the database by deallocating the underlying handle.

func (*RocksDB) Closed

func (r *RocksDB) Closed() bool

Closed returns true if the engine is closed.

func (*RocksDB) Compact

func (r *RocksDB) Compact() error

Compact forces compaction on the database.

func (*RocksDB) Destroy

func (r *RocksDB) Destroy() error

Destroy destroys the underlying filesystem data associated with the database.

func (*RocksDB) Flush

func (r *RocksDB) Flush() error

Flush causes RocksDB to write all in-memory data to disk immediately.

func (*RocksDB) Get

func (r *RocksDB) Get(key MVCCKey) ([]byte, error)

Get returns the value for the given key.

func (*RocksDB) GetProto

func (r *RocksDB) GetProto(
	key MVCCKey, msg proto.Message,
) (ok bool, keyBytes, valBytes int64, err error)

GetProto fetches the value at the specified key and unmarshals it.

func (*RocksDB) GetSSTables

func (r *RocksDB) GetSSTables() SSTableInfos

GetSSTables retrieves metadata about this engine's live sstables.

func (*RocksDB) GetStats

func (r *RocksDB) GetStats() (*Stats, error)

GetStats retrieves stats from this engine's RocksDB instance and returns it in a new instance of Stats.

func (*RocksDB) GetTempDir

func (r *RocksDB) GetTempDir() string

GetTempDir returns a temp path (usually under the store directory).

func (*RocksDB) Iterate

func (r *RocksDB) Iterate(start, end MVCCKey, f func(MVCCKeyValue) (bool, error)) error

Iterate iterates from start to end keys, invoking f on each key/value pair. See engine.Iterate for details.

func (*RocksDB) Merge

func (r *RocksDB) Merge(key MVCCKey, value []byte) error

Merge implements the RocksDB merge operator using the function goMergeInit to initialize missing values and goMerge to merge the old and the given value into a new value, which is then stored under key. Currently 64-bit counter logic is implemented. See the documentation of goMerge and goMergeInit for details.

The key and value byte slices may be reused safely. merge takes a copy of them before returning.

func (*RocksDB) NewBatch

func (r *RocksDB) NewBatch() Batch

NewBatch returns a new batch wrapping this rocksdb engine.

func (*RocksDB) NewIterator

func (r *RocksDB) NewIterator(prefix bool) Iterator

NewIterator returns an iterator over this rocksdb engine.

func (*RocksDB) NewSnapshot

func (r *RocksDB) NewSnapshot() Reader

NewSnapshot creates a snapshot handle from engine and returns a read-only rocksDBSnapshot engine.

func (*RocksDB) NewWriteOnlyBatch

func (r *RocksDB) NewWriteOnlyBatch() Batch

NewWriteOnlyBatch returns a new write-only batch wrapping this rocksdb engine.

func (*RocksDB) Put

func (r *RocksDB) Put(key MVCCKey, value []byte) error

Put sets the given key to the value provided.

The key and value byte slices may be reused safely. put takes a copy of them before returning.

func (*RocksDB) SetTempDir

func (r *RocksDB) SetTempDir(d string) error

SetTempDir allows overriding the tempdir returned by GetTempDir.

func (*RocksDB) String

func (r *RocksDB) String() string

String formatter.

type RocksDBBatchBuilder

type RocksDBBatchBuilder struct {
	// contains filtered or unexported fields
}

RocksDBBatchBuilder is used to construct the RocksDB batch representation. From the RocksDB code, the representation of a batch is:

WriteBatch::rep_ :=
   sequence: fixed64
   count: fixed32
   data: record[count]
record :=
   kTypeValue varstring varstring
   kTypeDeletion varstring
   kTypeSingleDeletion varstring
   kTypeMerge varstring varstring
   kTypeColumnFamilyValue varint32 varstring varstring
   kTypeColumnFamilyDeletion varint32 varstring varstring
   kTypeColumnFamilySingleDeletion varint32 varstring varstring
   kTypeColumnFamilyMerge varint32 varstring varstring
varstring :=
   len: varint32
   data: uint8[len]

The RocksDBBatchBuilder code currently only supports kTypeValue (BatchTypeValue), kTypeDeletion (BatchTypeDeletion)and kTypeMerge (BatchTypeMerge) operations. Before a batch is written to the RocksDB write-ahead-log, the sequence number is 0. The "fixed32" format is little endian.

The keys encoded into the batch are MVCC keys: a string key with a timestamp suffix. MVCC keys are encoded as:

<key>[<wall_time>[<logical>]]<#timestamp-bytes>

The <wall_time> and <logical> portions of the key are encoded as 64 and 32-bit big-endian integers. A custom RocksDB comparator is used to maintain the desired ordering as these keys do not sort lexicographically correctly. Note that the encoding of these keys needs to match up with the encoding in rocksdb/db.cc:EncodeKey().

func (*RocksDBBatchBuilder) Clear

func (b *RocksDBBatchBuilder) Clear(key MVCCKey)

Clear removes the item from the db with the given key.

func (*RocksDBBatchBuilder) Finish

func (b *RocksDBBatchBuilder) Finish() []byte

Finish returns the constructed batch representation. After calling Finish, the builder may be used to construct another batch, but the returned []byte is only valid until the next builder method is called.

func (*RocksDBBatchBuilder) Len

func (b *RocksDBBatchBuilder) Len() int

Len returns the number of bytes currently in the under construction repr.

func (*RocksDBBatchBuilder) Merge

func (b *RocksDBBatchBuilder) Merge(key MVCCKey, value []byte)

Merge is a high-performance write operation used for values which are accumulated over several writes. Multiple values can be merged sequentially into a single key; a subsequent read will return a "merged" value which is computed from the original merged values.

func (*RocksDBBatchBuilder) Put

func (b *RocksDBBatchBuilder) Put(key MVCCKey, value []byte)

Put sets the given key to the value provided.

type RocksDBBatchReader

type RocksDBBatchReader struct {
	// contains filtered or unexported fields
}

RocksDBBatchReader is used to iterate the entries in a RocksDB batch representation.

Example: r, err := NewRocksDBBatchReader(...)

if err != nil {
  return err
}
for r.Next() {
	 switch r.BatchType() {
	 case BatchTypeDeletion:
	   fmt.Printf("delete(%x)", r.UnsafeKey())
	 case BatchTypeValue:
	   fmt.Printf("put(%x,%x)", r.UnsafeKey(), r.UnsafeValue())
	 case BatchTypeMerge:
	   fmt.Printf("merge(%x,%x)", r.UnsafeKey(), r.UnsafeValue())
	 }
}
if err != nil {
  return nil
}

func NewRocksDBBatchReader

func NewRocksDBBatchReader(repr []byte) (*RocksDBBatchReader, error)

NewRocksDBBatchReader creates a RocksDBBatchReader from the given repr and verifies the header.

func (*RocksDBBatchReader) BatchType

func (r *RocksDBBatchReader) BatchType() BatchType

BatchType returns the type of the current batch entry.

func (*RocksDBBatchReader) Count

func (r *RocksDBBatchReader) Count() int

Count returns the declared number of entries in the batch.

func (*RocksDBBatchReader) Error

func (r *RocksDBBatchReader) Error() error

Error returns the error, if any, which the iterator encountered.

func (*RocksDBBatchReader) Next

func (r *RocksDBBatchReader) Next() bool

Next advances to the next entry in the batch, returning false when the batch is empty.

func (*RocksDBBatchReader) UnsafeKey

func (r *RocksDBBatchReader) UnsafeKey() []byte

UnsafeKey returns the key of the current batch entry. The memory is invalidated on the next call to Next.

func (*RocksDBBatchReader) UnsafeValue

func (r *RocksDBBatchReader) UnsafeValue() []byte

UnsafeValue returns the value of the current batch entry. The memory is invalidated on the next call to Next. UnsafeValue panics if the BatchType is BatchTypeDeleted.

type RocksDBCache

type RocksDBCache struct {
	// contains filtered or unexported fields
}

RocksDBCache is a wrapper around C.DBCache

func NewRocksDBCache

func NewRocksDBCache(cacheSize int64) RocksDBCache

NewRocksDBCache creates a new cache of the specified size. Note that the cache is refcounted internally and starts out with a refcount of one (i.e. Release() should be called after having used the cache).

func (RocksDBCache) Release

func (c RocksDBCache) Release()

Release releases the cache. Note that the cache will continue to be used until all of the RocksDB engines it was attached to have been closed, and that RocksDB engines which use it auto-release when they close.

type RocksDBSstFileReader

type RocksDBSstFileReader struct {
	// contains filtered or unexported fields
}

RocksDBSstFileReader allows iteration over a number of non-overlapping sstables exported by `RocksDBSstFileWriter`.

func MakeRocksDBSstFileReader

func MakeRocksDBSstFileReader(tempdir string) (RocksDBSstFileReader, error)

MakeRocksDBSstFileReader creates a RocksDBSstFileReader that uses a scratch directory which is cleaned up by `Close`.

func (*RocksDBSstFileReader) AddFile

func (fr *RocksDBSstFileReader) AddFile(path string) error

AddFile links the file at the given path into a database. See the RocksDB documentation on `AddFile` for the various restrictions on what can be added.

func (*RocksDBSstFileReader) Close

func (fr *RocksDBSstFileReader) Close()

Close finishes the reader.

func (*RocksDBSstFileReader) Iterate

func (fr *RocksDBSstFileReader) Iterate(
	start, end MVCCKey, f func(MVCCKeyValue) (bool, error),
) error

Iterate iterates over the keys between start inclusive and end exclusive, invoking f() on each key/value pair.

func (*RocksDBSstFileReader) NewIterator

func (fr *RocksDBSstFileReader) NewIterator(prefix bool) Iterator

NewIterator returns an iterator over this sst reader.

type RocksDBSstFileWriter

type RocksDBSstFileWriter struct {

	// DataSize tracks the total key and value bytes added so far.
	DataSize int64
	// contains filtered or unexported fields
}

RocksDBSstFileWriter creates a file suitable for importing with RocksDBSstFileReader.

func MakeRocksDBSstFileWriter

func MakeRocksDBSstFileWriter() RocksDBSstFileWriter

MakeRocksDBSstFileWriter creates a new RocksDBSstFileWriter with the default configuration.

func (*RocksDBSstFileWriter) Add

Add puts a kv entry into the sstable being built. An error is returned if it is not greater than any previously added entry (according to the comparator configured during writer creation). `Open` must have been called. `Close` cannot have been called.

func (*RocksDBSstFileWriter) Close

func (fw *RocksDBSstFileWriter) Close() error

Close finishes the writer, flushing any remaining writes to disk. At least one kv entry must have been added. Close is idempotent.

func (*RocksDBSstFileWriter) Open

func (fw *RocksDBSstFileWriter) Open(path string) error

Open creates a file at the given path for output of an sstable.

type SSTableInfo

type SSTableInfo struct {
	Level int
	Size  int64
	Start MVCCKey
	End   MVCCKey
}

SSTableInfo contains metadata about a single RocksDB sstable. This mirrors the C.DBSSTable struct contents.

type SSTableInfos

type SSTableInfos []SSTableInfo

SSTableInfos is a slice of SSTableInfo structures.

func (SSTableInfos) Len

func (s SSTableInfos) Len() int

func (SSTableInfos) Less

func (s SSTableInfos) Less(i, j int) bool

func (SSTableInfos) ReadAmplification

func (s SSTableInfos) ReadAmplification() int

ReadAmplification returns RocksDB's read amplification, which is the number of level-0 sstables plus the number of levels, other than level 0, with at least one sstable.

This definition comes from here: https://github.com/facebook/rocksdb/wiki/RocksDB-Tuning-Guide#level-style-compaction

func (SSTableInfos) String

func (s SSTableInfos) String() string

func (SSTableInfos) Swap

func (s SSTableInfos) Swap(i, j int)

type Stats

type Stats struct {
	BlockCacheHits           int64
	BlockCacheMisses         int64
	BlockCacheUsage          int64
	BlockCachePinnedUsage    int64
	BloomFilterPrefixChecked int64
	BloomFilterPrefixUseful  int64
	MemtableHits             int64
	MemtableMisses           int64
	MemtableTotalSize        int64
	Flushes                  int64
	Compactions              int64
	TableReadersMemEstimate  int64
}

Stats is a set of RocksDB stats. These are all described in RocksDB

Currently, we collect stats from the following sources:

  1. RocksDB's internal "tickers" (i.e. counters). They're defined in rocksdb/statistics.h
  2. DBEventListener, which implements RocksDB's EventListener interface.
  3. rocksdb::DB::GetProperty().

This is a good resource describing RocksDB's memory-related stats: https://github.com/facebook/rocksdb/wiki/Memory-usage-in-RocksDB

type Version

type Version struct {
	Version storageVersion
}

Version stores all the version information for all stores and is used as the format for the version file.

type Writer

type Writer interface {
	// ApplyBatchRepr atomically applies a set of batched updates. Created by
	// calling Repr() on a batch. Using this method is equivalent to constructing
	// and committing a batch whose Repr() equals repr. If sync is true, the
	// batch is synchronously written to disk. It is an error to specify
	// sync=true if the Writer is a Batch.
	ApplyBatchRepr(repr []byte, sync bool) error
	// Clear removes the item from the db with the given key.
	// Note that clear actually removes entries from the storage
	// engine, rather than inserting tombstones.
	Clear(key MVCCKey) error
	// ClearRange removes a set of entries, from start (inclusive) to end
	// (exclusive). Similar to Clear, this method actually removes entries from
	// the storage engine.
	ClearRange(start, end MVCCKey) error
	// ClearIterRange removes a set of entries, from start (inclusive) to end
	// (exclusive). Similar to Clear and ClearRange, this method actually removes
	// entries from the storage engine. Unlike ClearRange, the entries to remove
	// are determined by iterating over iter and per-key tombstones are
	// generated.
	ClearIterRange(iter Iterator, start, end MVCCKey) error
	// Merge is a high-performance write operation used for values which are
	// accumulated over several writes. Multiple values can be merged
	// sequentially into a single key; a subsequent read will return a "merged"
	// value which is computed from the original merged values.
	//
	// Merge currently provides specialized behavior for three data types:
	// integers, byte slices, and time series observations. Merged integers are
	// summed, acting as a high-performance accumulator.  Byte slices are simply
	// concatenated in the order they are merged. Time series observations
	// (stored as byte slices with a special tag on the roachpb.Value) are
	// combined with specialized logic beyond that of simple byte slices.
	//
	// The logic for merges is written in db.cc in order to be compatible with RocksDB.
	Merge(key MVCCKey, value []byte) error
	// Put sets the given key to the value provided.
	Put(key MVCCKey, value []byte) error
}

Writer is the write interface to an engine's data.

Directories

Path Synopsis
Package enginepb is a generated protocol buffer package.
Package enginepb is a generated protocol buffer package.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL