Scylla Compaction Strategies

Size tiered compaction strategy [STCS]

• STCS organises SSTable into tiers
• the tiers are based on the size of SSTable on an exponential scale
• When compacting several SSTables, a single SSTable is created.
  • It may be as large as union of all of them, then it's moved to next tier
  • or become much smaller due to deletes and expirations, potentially dropping to lower tier

STCS Space Amplification

• STCS requires space of at least twice the data size, this is called Space Amplification
• Temporary spaces due to compaction and
• Accumulation of updates and deletes across different tiers.

Levelled Compaction Strategy [LCS]

• This compaction is triggered when a level has more than 10 SSTables
• LCS picks one SSTable from level 'i', with size 'x', to compact
• it then roughly finds 10 SSTables in the next level
  • Overlapping with SSTable and compacts all of them together
• It then writes the resulting run to next level, run size bound by (1+10)*X
• while LCS limits space amplification but it results in higher write amplification

Time Window Compaction Strategy [TWCS]

• Memtables have a write time, SSTables inherit this write time
• Only SSTables that belong to the same window are compacted together

create table twcs.example (
     id int,
     value int,
     text_value text,
     PRIMARY KEY (id, value)
) with clustering order by (value ASC)
       AND compaction = {
           'compaction_window_size' : '1',
               'compaction_window_unit' : 'DAYS', 
               'class' : TimeWIndowCompactionStrategy', 
       } ;

• All data in the partition is inserted in the same window or in small number of windows
• Deletes and writes -
  • Data was written 1 year ago, data point is in SSTable 1
  • Delete will be written now, tombstone is in the SSTable N
  • Reads have to read them both
  • Data is never really gone

Incremental Compaction Strategy

• Observed problems with legacy compaction Strategies
• STCS and LCS has high space and write Amplification respectively.
• Sorted set of SSTables
• The SSTables are non-overlapping
  • those are called Fragments
• A run is equivalent to a large SSTable, split into several smaller SSTables
• Fragments are disjoint and sorted with respect to each other, so we scan the runs, fragment by fragment and compact them increamentally