CASSANDRA READ/WRITE PATH Josh McKenzie

josh.mckenzie@datastax.com

CORE COMPONENTS

Core Components

Memtable data in memory (R/W) CommitLog data on disk (W/O) SSTable data on disk (immutable, R/O) CacheService (Row Cache and Key Cache) in-memory caches ColumnFamilyStore logical grouping of table data

DataTracker and View provides atomicity and grouping of memtable/sstable data

ColumnFamily Collection of Cells (Sorted map of Columns) Cell Name, Value, TS Tombstone Deletion marker indicating TS and deleted cell(s)

MemTable

In-memory data structure consisting of: Memory pools (on-heap, off-heap) Allocators for each pool Size and limit tracking and CommitLog sentinels

Map of Key ! AtomicBTreeColumns Atomic copy-on-write semantics for row-data Flush to disk logic is triggered when pool passes ratio of usage relative

to user-configurable threshold Memtable w/largest ratio of used space (either on or off heap) is flushed

to disk

On heap vs. Off heap Memtables: an overview

http://www.datastax.com/dev/blog/off-heap-memtables-in-cassandra-2-1 https://issues.apache.org/jira/browse/CASSANDRA-6689 https://issues.apache.org/jira/browse/CASSANDRA-6694 memtable_allocation_type

offheap_buffers moves the cell name and value to DirectBuffer objects. The values are still live Java buffers. This mode only reduces heap significantly when you are storing large strings or blobs

offheap_objects moves the entire cell off heap, leaving only the NativeCell reference containing a pointer to the native (off-heap) data. This makes it effective for small values like ints or uuids as well, at the cost of having to copy it back on-heap temporarily when reading from it.

Default in 2.1 is heap buffers

On heap vs. Off heap: continued

Why? Reduces sizes of objects in memory no more ByteBuffer overhead More data fitting in memory == better performance

Code changes that support it: MemtablePools allow on vs. off-heap allocation (and Slab, for that matter) MemtableAllocators to allow differentiating between on-heap and off-heap

allocation

DecoratedKey and *Cells changed to interfaces to have different allocation implementations based on native vs. heap

CommitLog

Append-only file structure corresponding provides interim durability for writes while theyre living in Memtables and havent been flushed to sstables

Has sync logic to determine the level of durability to disk you want - either PeriodicCommitLogService or BatchCommitLogService Periodic: (default) checks to see if it hit window limit, if so, block and wait for sync to catch up Batch: no ack until fsync to disk. Waits for a specific window before hitting fsync to coalesce

Singleton faade for commit log operations Consists of multiple components

CommitLog.java: interface to subsystem CommitLogSegment.java: files on disk CommitLogManager.java: segment allocation and management CommitLogArchiver.java: user-defined commands pre/post flush CommitLogMetrics.java

SSTable

Ordered-map of KVP Immutable Consist of 3 files: Bloom Filter: optimization to determine if the Partition Key youre

looking for is (probably) in this sstable

Index file: contains offset into data file, generally memory mapped Data file: contains data, generally compressed

Read by SSTableReader

CacheService.java

In-memory caching service to optimize lookups of hot data Contains three caches:

keyCache rowCache counterCache

See: AutoSavingCache.java InstrumentingCache.java

Tunable per table, limits in cassandra.yaml, keys to cache, size in mb, rows, size in mb Defaults to keys only, can enable row cache via CQL

ColumnFamilyStore.java

Contains logic for a table Holds DataTracker Creating and removing sstables on disk Writing / reading data Cache initialization Secondary index(es) Flushing memtables to sstables Snapshots And much more (just short of 3k LoC)

CFS: DataTracker and View

DataTracker allows for atomic operations on a view of a Table (ColumnFamilyStore) Contains various logic surrounding Memtables and flushing, SSTables and

compaction, and notification for subscribers on changes to SSTableReaders

1 DataTracker per CFS, 1 AtomicReference per DataTracker View consists of current Memtable, Memtables pending flush, SSTables for the CFS,

and SSTables being actively compacted

Currently active Memtable is atomically switched out in: DataTracker.switchMemtable(boolean truncating)

ColumnFamily.java

A sorted map of columns Abstract class, extended by:

ArrayBackedSortedColumns Array backed Non-thread-safe Good for iteration, adding cells (especially if in sorted order)

AtomicBTreeColumns (memtable only) Btree backed Thread-safe w/atomic CAS Logarithmic complexity on operations

Logic to add / retrieve columns, counters, tombstones, atoms

THE READ PATH

Read Path: Very High Level

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

Read-specific primitive: QueryFilter

Contains IDiskAtomFilter IDiskAtomFilter: used to get columns from Memtable, SSTable, or SuperColumn

IdentityQueryFilter, NamesQueryFilter, SliceQueryFilter

Contains a variety of iterators to collate on disk contents, gather tombstones, reduce (merge) Cells with the same name, etc

See: collateColumns() gatherTombstones() getReducer(final Comparator comparator)

Read-specific class: SSTableReader

Has 2 SegmentedFiles, ifile and dfile, for index and data respectively Contains a Key Cache, caching positions of keys in the SSTR Contains an IndexSummary w/sampling of the keys that are in the table

Binary search used to narrow down location in file via IndexSummary getIndexScanPosition(RowPosition key)

Short running operations guarded by ColumnFamilyStore.readOrdering See OpOrder.java producer/consumer synchronization primitive to coordinate readers w/

flush operations

Resources are ref counted see RefCounted.java, Tidy interface, various private classes in SSTR that implement Tidy interface to clean up resources

Provides methods to retrieve an SSTableScanner which gives you access to OnDiskAtoms via iterators and holds RandomAccessReaders on the raw files on disk OnDiskAtom ! Cell ! *Cell

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

ReadVerbHandler and ReadCommands

Messages are received by the MessagingService and passed to the ReadVerbHandler for appropriate verbs

ReadCommands: SliceFromReadCommand

Relies on SliceQueryFilter, uses a range of columns defined by a ColumnSlice SliceByNamesReadCommand

Relies on NamesQueryFilter, uses a column name to retrieve a single column Both diverge in calls and converge back into implementers of ColumnFamily

ArrayBackedSortedColumns, AtomicBTreeSortedColumns

public Row Keyspace.getRow(QueryFilter filter) {

ColumnFamilyStore cfStore = getColumnFamilyStore(filter.getColumnFamilyName());

ColumnFamily columnFamily = cfStore.getColumnFamily(filter);

return new Row(filter.key, columnFamily);

}

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

RowCache

CFS.getThroughCache(UUID cfId, QueryFilter filter) After retrieving our CFS, the first thing we check is our Row Cache to see if the row is

already merged, in memory, and ready to go

If we get a cache hit on the key, well: Confirm its not just a sentinel of someone else in flight. If so, we query w/out caching If the data for the key is valid, we filter it down to the query we have in flight and return

those results as itll have >= the count of Cells were looking for

On cache miss: Eventually cache all top level columns for the key queried if configured to do so (after

Collation) Cache results of user query if it satisfies the cache config params Extend the results of the query to satisfy the caching requirements of the system

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

CollationController.collect*Data ()

The data were looking for may be in a Memtable, an SSTable, multiple of either, or a combination of all of them.

The logic to query this data and merge our results exists in CollationController.java: collectAllData collectTimeOrderedData

High level flow: 1. Get data from memtables for the QueryFilter were processing 2. Get data from sstables for the QueryFilter were processing 3. Merge all the data together, keeping the most recent 4. If we iterated across enough sstables, hoist up the now defragmented data into a memtable,

bypassing CommitLog and Index update (collectTimeOrderedData only)

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

CollationController merging: memtables

Fairly straightforward operations on memtables in the view: Check all memtables to see if they have a ColumnFamily that matches our filter.key Add all columns to our result ColumnFamily that match Keep a running tally of the mostRecentRowTombstone for use in next step.

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

CollationController merging: sstables

We have a few optimizations available for merging in data from sstables: Sort the collection of SSTables by the max timestamp present Iterate across the SSTables

Skipping any that are older than the most recent tombstone weve seen Create a reduced name filter by removing columns from our filter where we

have fresher data than the SSTRs max Timestamp

Get iterator from SSTR for Atoms matching that reduced name filter Add any matching OnDiskAtoms to our result set (BloomFilter excludes via

iterator with SSTR.getPosition() call)

Overview the Read Path

Return results

CollationController

Keyspace

ColumnFamilyStore

Check Row Cache hit

miss

Memtable

SSTables

read merge

ColumnFamily

Update Row Cache

Coordinator

MessagingService

Key Cache

Binary scan index, update cache

Seek to cached position

hit

miss

THE WRITE PATH

Write Path: Very High Level

Overview the Write Path

MessagingService

Keyspace

CommitLog Enabled for this mutation?

Yes

Write CommitLog

No

Skip

Write to Memtable

SecondaryIndexManager.Updater

Invalidate Row Cache

MutationVerbHandler, Mutation.apply

Contains Keyspace name DecoratedKey Map of cfId to ColumnFamily of modifications to perform MutationVerbHandler ! Mutation.apply() ! Keyspace.apply() !

ColumnFamilyStore.apply()

Overview the Write Path

MessagingService

Keyspace

CommitLog Enabled for this mutation?

Yes

Write CommitLog

No

Skip

Write to Memtable

SecondaryIndexManager.Updater

Invalidate Row Cache

The CommitLog ecosystem

CommitLogSegment: file on disk CommitLogSegmentManager: allocation and recycling of CommitLogSegments CommitLogArchiver: allows user-defined archive and restore commands to be run

Reference conf/commitlog_archiving.properties An AbstractCommitLogService, one of either:

BatchCommitLogService writer waits on sync to complete before returning PeriodicCommitLogService Check if sync is behind, if so, register w/signal and

block until lastSyncedAt catches up

CommitLogSegmentManager (CLSM): overview

Contains 2 collections of CommitLogSegments availableSegments: Segments ready to be used activeSegments: Segments that are active and contain unflushed data

Only 1 active CommitLogSegment is in use at any given time Manager thread is responsible for maintaining active vs. available

CommitLogSegments and can be woken up by other contexts when maintenance is needed

CLSM: allocation on the write path

During CommitLog.add(), a writer asks for allocated space for their mutation from the CommitLogSegmentManager

This is passed to the active CommitLogSegments allocate() method CommitLogSegmentManager.allocate(int size) spins non-blocking until the space in

the segment is allocated, at which time it marks it dirty

If the CLS.allocate() call returns null indicating we need a new segment: CommitLogSegment.advanceAllocatingFrom(CommitLogSegment old) Goal is to move CLS from available to active segments so we have more CLS to work with If it fails to get an available segment, the manager thread is woken back up to do some

maintenance, be it recycling or allocating a new CLS

CLSM: manager thread, new segments, recycling

Constructor creates a runnable that blocks on segmentManagementTasks Task can either be null indicating were out of space (allocate path) or a segment thats

flushed and ready for recycle

If theres no available segments, we create new CommitLogSegments and add them to availableSegments hasAvailableSegments WaitQueue is signaled by this to awake any blocking writes waiting for

allocation

When our CommitLog usage is approaching our allowable limit: If our total used size is > than the size allowed

CommitLogSegmentManager.flushDataFrom on a list of activeSegments Force flush on any CFS thats dirty

Which switches Memtables and flushes to SSTable more on this later

Overview the Write Path

MessagingService

Keyspace

CommitLog Enabled for this mutation?

Yes

Write CommitLog

No

Skip

Write to Memtable

SecondaryIndexManager.Updater

Invalidate Row Cache

Memtable writes

We attempt to get the partition for the given key if it exists If not, we allocate space for a new key and put an empty entry in the memtable for it,

backing that out if we race and someone else got there first on allocation

Once we have space allocated, we call addAllWithSizeDelta Add the record to a new BTree and CAS it into the existing Holder Updates secondary indexes Finalize some heap tracking in the ColumnUpdater used by the BTree to perform updates

Further reading: AtomicBTreeColumns.java (specifically addAllWithSizeDelta) BTree.java

MemtablePool

Single MEMORY_POOL instance across entire DB Get an allocator to the memory pool during construction of a memtable Interface covering management of an on-heap and off-heap pool via SubPool

HeapPool: On heap ByteBuffer allocations and release, subject to GC w/object overhead NativePool: Blend of on and off heap based on limits passed in

Off heap allocations and release through NativeAllocator, calls to Unsafe SlabPool: Blend of on and off heap based on limits passed in (used for memtables, reduced fragmentation)

Allocated in large chunks by SlabAllocator (1024*1024)

MemtablePool.SubPool / SubAllocator: Contains various atomically updated longs tracking:

Limits on allocation Currently allocated amounts Currently reclaiming amounts Threshold for when to run Cleaner thread

Spin and CAS for updates on the above on allocator calls in addAllWithSizeDelta

Overview the Write Path

MessagingService

Keyspace

CommitLog Enabled for this mutation?

Yes

Write CommitLog

No

Skip

Write to Memtable

SecondaryIndexManager.Updater

Invalidate Row Cache

Secondary Indexes: an overview

Essentially a separate table stored on disk / in memtable Contains a ConcurrentNavigableMap of ByteBuffer ! SecondaryIndex

There are quite a few SecondaryIndex implementations in the code base, ex: PerRowSecondaryIndex PerColumnSecondaryIndex KeysIndex

On Write Path: SecondaryIndex updater passed down through to ColumnUpdater ctor On ColumnUpdater.apply(), insert for secondary index is called Essentially amounts to a 2nd write on another table

Overview the Write Path

MessagingService

Keyspace

CommitLog Enabled for this mutation?

Yes

Write CommitLog

No

Skip

Write to Memtable

SecondaryIndexManager.Updater

Invalidate Row Cache

FLUSHING MEMTABLES

CLSM.activeSegments ColumnFamilyStore

Flushing Memtables

Memtable

SSTableWriter

SSTable

SSTableReader

CommitLog.discardCompletedSegments( cfId, lastReplayPosition)

CLS 1

CLS 2

CLS Active Actively allocating

Skip

Still other cfDirty Remove flushed cfId

Removed last dirty Recycle CLS

Stop at position of flush

MemtableCleanerThread: starting a flush

When MemtableAllocator adjusts the size of the data it has acquired the MemtablePool checks whether or not we need to flush to free up space in memory

If our used memory is > than the total reclaiming memory + the limit * ratio defined in conf.memtable_cleanup_threshold, a memtable needs to be cleaned

Cleaner thread is currently: ColumnFamilyStore.FlushLargestColumnFamily()) We find the memtable with the largest Ownership ratio as determined by the currently

owned memory vs. limit, taking the max of either on or off heap

Signals to CommitLog to discard completed segments on PostFlush stage of flush

Memtable Flushing

Reference ColumnFamilyStore$Flush 1st, switch out memtables in CFS.DataTracker.View so new ops go to new memtable Sets lifecycle in memtable to discarding Runs the FlushRunnable in the Memtable

Memtable.writeSortedContents Uses SSTableWriter to write sorted contents to disk Returns SSTableReader created by SSTableWriter.closeAndOpenReader

Memtable.setDiscarded() ! MemtableAllocator.setDiscarded() Lifecycle to Discarded Free up all memory from the allocator for this memtable

Memtable Flushing: the commit log

ColumnFamilyStore$PostFlush All relative to a timestamp of the most recent data in the flushed memtable Record sentinel for when this cf was cleaned (to be used later if it was active and we

couldnt purge at time of flush)

Walk through CommitLogSegments and remove dirty cfid Unless its actively being allocated from If the CLS is no longer in use:

Remove it from our activeSegments Queue a task for Management thread to wake up and recycle the segment

Switching out memtables

CFS.switchMemtableIfCurrent / CFS.switchMemtable Theres some complex non-blocking write-barrier operations on

Keyspace.writeOrder to allow us to wait for writes to finish in this context before swapping out with new memtables regardless of dirty status

Reference: OpOrder.java,OpOrder.Barrier

Write sorted contents to disk (Memtable.FlushRunnable.runWith(File sstableDirectory)

cfs.replaceFlushed, swapping the memtable with the new SSTableReader returned from writeSortedContents

3.0

A couple of interesting things

Sylvain has a herculean effort underway to refactor and modernize the storage engine, removing quite a bit of complexity from the querying process

Various components renamed (names are in flux atm) https://issues.apache.org/jira/browse/CASSANDRA-8099

CommitLogSegment Recycling is pretty complex without much payoff, so theres plans to remove it:

https://issues.apache.org/jira/browse/CASSANDRA-8771