• Julien Muchembled's avatar
    storage: optimize storage layout of raw data for replication · f4dd4bab
    Julien Muchembled authored
    # Previous status
    
    The issue was that we had extreme storage fragmentation from the point of view
    of the replication algorithm, which processes one partition at a time.
    
    By using an autoincrement for the 'data' table, rows were ordered by the time
    at which they were added:
    - parts may be the result of replication -> ordered by partition, tid, oid
    - other rows are globally sorted by tid
    
    Which means that when scanning a given partition, many rows were skipped all
    the time:
    - if readahead is bigger enough, the efficiency is 1/N for a node with N
      partitions assigned
    - else, it is worse because it seeks all the time
    
    For huge databases, the replication was horribly slow, in particular from HDD.
    
    # Chosen solution
    
    This commit changes how ids are generated to somehow split 'data'
    per partition. The backend tracks 1 last id per assigned partition, where the
    16 higher bits contains the partition. Keep in mind that the value of id has no
    meaning and it's only chosen for performance reasons. IOW, a row can be
    referred by an oid of a partition different than the 16 higher bits of id:
    - there's no migration needed and the 16 higher bits of all existing rows are 0
    - in case of deduplication, a row can still be shared by different partitions
    
    Due to https://jira.mariadb.org/browse/MDEV-12836, we leave the autoincrement
    on existing databases.
    
    ## Downsides
    
    On insertion, increasing the number of partitions now slows down significantly:
    for 2 nodes using TokuDB, 4% for 180 partitions, 40% for 2000. For 12
    partitions, the difference remains negligible. The solution for this issue will
    be to enable to increase the number of partitions efficiently, so that nodes
    can keep a small number of them, even for DB that are expected to grow so much
    that many nodes are added over time: such feature was already considered so
    that users don't have to worry anymore about this obscure setting at database
    creation.
    
    Read performance is only slowed down for applications that read a lot of data
    that were written contiguously, but split in small blocks. A solution is to
    extend ZODB so that the application tells it to chose new oids that will end up
    in the same partition. Like for insertion, there should not be too many
    partitions.
    
    With RocksDB (MariaDB 10.2.10), it takes a significant amount of time to
    collect all last ids at startup when there are many partitions.
    
    ## Other advantages
    
    - The storage layout of data is now always the same and does not depend on
      whether rows came from replication or commits.
    - Efficient deletion of partition to free space in-place will be possible.
    
    # Considered alternative
    
    The only serious alternative was to replicate as many partitions as possible at
    the same time, ideally all assigned partitions, but it's not always possible.
    For best performance, it would often require to synchronize new nodes, or even
    all of them, so that thesource nodes don't have to scan 'data' several times.
    
    If existing nodes are kept, all data that aren't copied to the newly added
    nodes have to be skipped. If the number of nodes is multiplied by N, the
    efficiency is 1-1/N at best (synchronized nodes), else it's even worse
    because partitions are somehow shuffled.
    
    Checking/replacing a single node would remain slow when there are several
    source nodes.
    
    At last, such an algorithm would be much more complex and we would not have the
    other advantages listed above.
    f4dd4bab
manager.py 30.4 KB