scylladb/docs/architecture/tablets.rst

=========================================
Data Distribution with Tablets
=========================================

A ScyllaDB cluster is a group of interconnected nodes. The data of the entire 
cluster has to be distributed as evenly as possible across those nodes.

ScyllaDB is designed to ensure a balanced distribution of data by storing data
in tablets. When you add or remove nodes to scale your cluster, add or remove
a datacenter, or replace a node, tablets are moved between the nodes to keep
the same number on each node. In addition, tablets are balanced across shards
in each node.

This article explains the concept of tablets and how they let you scale your
cluster quickly and seamlessly.

Data Distribution
-------------------

ScyllaDB distributes data by splitting tables into tablets. Each tablet has 
its replicas on different nodes, depending on the RF (replication factor). Each
partition of a table is mapped to a single tablet in a deterministic way. When you
query or update the data, ScyllaDB can quickly identify the tablet that stores
the relevant partition. 

The following example shows a 3-node cluster with a replication factor (RF) of
3. The data is stored in a table (Table 1) with two rows. Both rows are mapped
to one tablet (T1) with replicas on all three nodes.

.. image:: images/tablets-cluster.png

.. TODO - Add a section about tablet splitting when there are more triggers,
   like throughput. In 6.0, tablets only split when reaching a threshold size
   (the threshold is based on the average tablet data size).

Load Balancing
==================

ScyllaDB autonomously moves tablets to balance the load. This process
is managed by a load balancer mechanism and happens independently of
the administrator. The tablet load balancer decides where to migrate
the tablets, either within the same node to balance the shards or across 
the nodes to balance the global load in the cluster.

During this process, the balancer will compute disk usage based on the
actual disk sizes of the tablets located on the nodes. It will then issue
tablet migrations which migrate tablets from nodes with higher to nodes
with lower disk utilization, equalizing the load.

As a table grows, each tablet can split into two, creating a new tablet.
The load balancer can migrate the split halves independently to different nodes
or shards.

The load-balancing process takes place in the background and is performed
without any service interruption.

Scaling Out
=============

A tablet can be dynamically migrated to an existing node or a newly added
empty node. Paired with consistent topology updates with Raft, tablets allow
you to add multiple nodes simultaneously. After nodes are added to the cluster,
existing nodes stream data to the new ones, and the system load eventually
converges to an even distribution as the process completes. 

With tablets enabled, manual cleanup is not required.
Cleanup is performed automatically per tablet,
making tablets-based streaming user-independent and safer.

In addition, tablet cleanup is lightweight and efficient, as it doesn't
involve rewriting SStables on the existing nodes, which makes data ownership 
changes faster. This dramatically reduces 
the impact of cleanup on the performance of user queries.

The following diagrams show migrating tablets from heavily loaded nodes A and B
to a new node.

.. image:: images/tablets-load-balancing.png

File-based Streaming
========================

Migrating tablets is performed by streaming entire
SStables, which does not require (de)serializing or processing mutation fragments.
As a result, less data is streamed over the network, and less CPU is consumed,
especially for data models that contain small cells.

File-based streaming is used for tablet migration in all 
:ref:`keyspaces created with tablets enabled <tablets>`.

.. _absolute-number-of-tablets:

Absolute number of tablets
==========================

ScyllaDB has a background process that periodically re-evaluates the number of tablets of each table.
The computed number of tablets a table will have is based on several parameters and factors. These are:

* Keyspace tablets option ``'initial'``. This option sets the initial number of tablets on the keyspace level.
  See :ref:`The tablets property <tablets>` for details.
* Table-level option ``'expected_data_size_in_gb'``. This option sets the minimal number of tablets for a table
  based on the expected table size and the target tablet size. See
  :ref:`Per-table tablet options <cql-per-table-tablet-options>` for details.
* Table-level option ``'min_per_shard_tablet_count'``. Using this option results in the number of tablets being
  computed based on the number of shards in a DC so that each shard has at least ``'min_per_shard_tablet_count'``
  tablets on average. See :ref:`Per-table tablet options <cql-per-table-tablet-options>` for details.
* Table-level option ``'min_tablet_count'``. This option sets the minimal number of tablets for the given table.
  See :ref:`Per-table tablet options <cql-per-table-tablet-options>` for details.
* Config option ``'tablets_initial_scale_factor'``. This option sets the minimal number of tablets per shard
  per table globally. This option can be overridden by the table-level option: ``'min_per_shard_tablet_count'``.
  ``'tablets_initial_scale_factor'`` is ignored if either the keyspace option ``'initial'`` or table-level
  option ``'min_tablet_count'`` is set.

Another factor that determines the absolute tablet count is the amount of data the table contains. If the
amount of data in the table is such that the average tablet size is larger than double the target tablet size,
the table will be split (the number of tablets will be doubled), and if the average tablet size is smaller than
half the target tablet size, it will be merged (the number of tablets will be halved).

Each of these factors is taken into consideration, and the one producing the largest number of tablets wins, and
will be used as the number of tablets for the given table.

As the last step, in order to avoid having too many tablets per shard, which could potentially lead to overload
and performance degradation, ScyllaDB will run the following algorithm to respect the ``tablets_per_shard_goal``
config option:

* Compute average tablet count per-shard in each DC.
* Determine if per-shard goal is exceeded in that DC.
* Compute scale factor by which tablet count should be multiplied so that the goal is not exceeded in that DC.
* Take the smallest scale factor among all DCs, which ensures that no DC is overloaded.
* Each table's tablet count is aligned to the nearest power of 2 post-scaling.

Please note that because of this alignment, the scaling may not be effective and in the worst case may be
overshot by a factor of 2, and that the ``tablets_per_shard_goal`` is a soft limit and not a hard constraint.

Finally, the computed tablet count is compared with the current tablet count for each table, and if there is
a difference, a table resize (split or merge) is executed.

.. _tablets-enable-tablets: 

Enabling Tablets
-------------------

ScyllaDB now uses tablets by default for data distribution.
Enabling tablets by default when creating new keyspaces is
controlled by the :confval:`tablets_mode_for_new_keyspaces` option. However, tablets only work if
supported on all nodes within the cluster.

When creating a new keyspace with tablets enabled by default, you can still opt-out
on a per-keyspace basis using ``CREATE KEYSPACE <ks> WITH tablets = {'enabled': false}``,
unless the :confval:`tablets_mode_for_new_keyspaces` option is set to ``enforced``.

Note: The recommended ``NetworkTopologyStrategy`` for keyspaces
remains *required* even if tablets are disabled.

.. code:: cql

    CREATE KEYSPACE my_keyspace
    WITH replication = {
        'class': 'NetworkTopologyStrategy',
        'replication_factor': 3
    } AND tablets = {
        'enabled': false
    };

When creating a new keyspace with tablets disabled by default, you can still opt-in
on a per-keyspace basis. The recommended ``NetworkTopologyStrategy`` for keyspaces
remains *required* when using tablets.

You can create a keyspace with tablets enabled with the ``tablets = {'enabled': true}`` option:

.. code:: cql

    CREATE KEYSPACE my_keyspace
    WITH replication = {
        'class': 'NetworkTopologyStrategy',
        'replication_factor': 3
    } AND tablets = {
        'enabled': true
    };


.. warning::

    You cannot ALTER a keyspace to enable or disable tablets.
    The only way to update the tablet support for a keyspace is to DROP it
    (losing the schema and data) and then recreate it after redefining 
    the keyspace schema with ``tablets = { 'enabled': false }`` or 
    ``tablets = { 'enabled': true }``.

.. _keyspace-rf-rack-valid-to-enforce-rack-list:

Enforcing Rack-List Replication for Tablet Keyspaces
------------------------------------------------------------------

The ``rf_rack_valid_keyspaces`` is a legacy option that ensures that all keyspaces with tablets enabled are
:term:`RF-rack-valid <RF-rack-valid keyspace>`.

Requiring every tablet keyspace to use the rack list replication factor exclusively is enough to guarantee the keyspace is
:term:`RF-rack-valid <RF-rack-valid keyspace>`. It reduces restrictions and provides stronger guarantees compared
to ``rf_rack_valid_keyspaces`` option.

To enforce rack list in tablet keyspaces, use ``enforce_rack_list`` option. It can be set only if all tablet keyspaces use
rack list. To ensure that, follow a procedure of :ref:`conversion to rack list replication factor <conversion-to-rack-list-rf>`.
After that restart all nodes in the cluster, with ``enforce_rack_list`` enabled and ``rf_rack_valid_keyspaces`` disabled. Make
sure to avoid setting or updating replication factor (with CREATE KEYSPACE or ALTER KEYSPACE) while nodes are being restarted.

.. _tablets-limitations:

Limitations and Unsupported Features
--------------------------------------

.. warning::

    If a keyspace has tablets enabled and contains a materialized view or a secondary index, it must
    remain :term:`RF-rack-valid <RF-rack-valid keyspace>` throughout its lifetime. Failing to keep that
    invariant satisfied may result in data inconsistencies, performance problems, or other issues.

    The invariant is enforced while the keyspace contains a materialized view or a secondary index, or
    if the `rf_rack_valid_keyspaces` option is set, by rejecting operations that would violate the RF-rack-valid property:
    Altering a keyspace's replication factor, adding a node in a new rack, or removing the last node
    in a rack, will be rejected if they would result in an RF-rack-invalid keyspace.

To enable materialized views and secondary indexes for tablet keyspaces, the keyspace
must be :term:`RF-rack-valid <RF-rack-valid keyspace>`.
See :ref:`Views with tablets <admin-views-with-tablets>` for details.

Resharding in keyspaces with tablets enabled has the following limitations:

* ScyllaDB does not support reducing the number of shards after node restart.
* ScyllaDB does not reshard data on node restart. Tablet replicas remain
  allocated to the old shards on restart and are subject to background
  load-balancing to additional shards after restart completes and the node 
  starts serving CQL.