#DBHangOps 10/30/14 -- New in TokuDB, Outbrain's Orchestrator, and more!
Hello everybody!
Join in #DBHangOps this Thursday, October, 30, 2014 at 11:00am pacific (18:00 GMT), to participate in the discussion about:
- What's new in TokuDB 7.5
- Outbrain's Orchestrator - how are other people using it?
- Mixing data and metadata
You can check out the event page at https://plus.google.com/events/c133qbpofubhliavlou8iie4qic on Thursday to participate.
As always, you can still watch the #DBHangOps twitter search, the @DBHangOps twitter feed, or this blog post to get a link for the google hangout on Thursday!
See all of you on Thursday!
You can catch a livestream at:
What's new in TokuDB 7.5
- What is TokuDB?
- Indexing 101!
- B-Trees
- InnoDB and MyISAM use BTree indexes. It typically looks like a pyramid with "Pivots" and points that can get you to the sorted leaf nodes.
- This allows you to quickly get to the data you're looking for.
- Imagine the data in the BTree index being simply Key=>Value pairs.
- Once you get to the leaf nodes, a binary search will happen, but everything is already sorted so it's a quick scan
- BTrees were invented in the early 1970s. A good data structure, but we're starting to push it's scale limits
- When the BTree doesn't wholly fit in memory, you'll start to see "problems".
- One IO is needed for every insert/update (actually it's one IO for every index on the table)
- Fractal trees
- similar to B-Trees. The 2 big differences are:
- All internal nodes have message buffers
- Big nodes (4MB vs. ~16KB)
- You can define the node size at a table level
- thanks to the message buffers, you can sometimes serve reads before traversing farther down the index. Eventually changes cascade down into the leaf nodes.
- Updates are effectively a "Delete then Insert". So only inserts/deletes were needed initially
- New message types allow for greater flexibility too.
- add_column() message means a new field can be kep in the message buffer before persisting down in the leaf nodes.
- What about the InnoDB Change buffer?
- The primary key index doesn't have a buffer
- Each secondary index only has one buffer (and it can grow pretty large)
- The bigger the dataset, the slower your insertions will get over time
- MySQL Replication 101
- Modes and logging:
- Statement Based - statement that's run on a master is written verbatim into the replication log
- Row Based - inserts and deletes simply write a PK and the operation. Updates write a copy of the row before the update and the changed version of the row
- Mixed - defaults to statement unless MySQL detects an "unsafe" statement (e.g. a non-deterministic UPDATE statement)
- Read only slaves
- typically you want to set "read_only=1". This means only replication or users with the SUPER privilege can change data
- "Read-free Replication" in TokuDB will break if a statement is inserted by a SUPER user. Be warned!
- Slave lag
- in MySQL 5.5, replication is single-threaded. The master is concurrent, but the slave will not be.
- MySQL 5.6 and MariaDB 10.0 have parallel replication methods
- Read Free Replication in TokuDB benefits from these too!
- TokuDB Read Free Replication
- Binlog format must be set to ROW
- Slave must have the following set:
- read_only=1
- tokudb_rpl_unique_checks=0 and/or tokudb_rpl_lookup_rows=0
- Skipping unique checks (tokudb_rpl_unique_checks)
- Master's already done the uniqueness check, so the slave can probably skip this
- Since InnoDB doesn't support change buffering on the primary key, it needs to do this uniqueness check for maintenance
- Skipping Read/Write/Modify behavior (tokudb_rpl_lookup_rows)
- If RBR is enabled, master is already providing before/after images of a row
- Everything that's needed for the change is already in the binary log
- TokuDB can do a simple message injection
- Enabling Read Free Replication can help lagging slaves keep up. It also opens up a significant percentage of read capacity.
- When enabling Read Free Replication , most reads on a slave that would be needed for replication effectively go away.
- By opening up read capacity, you could potentially co-locate more than one mysqld for different mysql clusters to save on hardware
Outbrain Orchestrator