Lightweight Transactions¶

The IF clause¶

Creating a conditional statement with an IF clause can include a query of individual cells and or collection elements. It can use IN clauses or comparison operators, such as <, >, >=, <=, ==, and !=. The important part to keep in mind is that in order to initiate the transaction, the condition must evaluate to true and if there are several different elements in the conditional statement, each part must evaluate to true. If any part evaluates to false the transaction does not complete.

The IF condition can be created from any of the following CQL components: IF EXISTS, IF NOT EXISTS, or one or more predicates on the existing row.

What row does IF clause apply to?¶

Conditional statements which evaluate or assign non-static columns must specify both the clustering key and the partition key. Such statements are said to apply to regular rows; statements which only restrict the partition key must use only static columns and are said to apply to the static row of the partition.

A regular row exists if at least one regular cell or the clustering key is assigned. For example, consider the following table:

CREATE TABLE t (
    p INT,
    c INT,
    r INT,
    s INT STATIC,
    PRIMARY KEY(p, c));

It has a partition key p, a clustering key c, a regular cell r and a static cell s.

To materialize a regular row, it’s sufficient to assign any of c or r:

> INSERT INTO t (p, c, r) VALUES (1,1,NULL) IF NOT EXISTS;
+-------------+------+------+------+------+
| [applied]   | p    | c    | s    | r    |
|-------------+------+------+------+------|
| True        | null | null | null | null |
+-------------+------+------+------+------+
> INSERT INTO t (p, c, r) VALUES (1,1,NULL) IF NOT EXISTS;
+-------------+-----+-----+------+------+
| [applied]   |   p |   c | s    | r    |
|-------------+-----+-----+------+------|
| False       |   1 |   1 | null | null |
+-------------+-----+-----+------+------+

To materialize a static row, one must explicitly assign at least one static cell: otherwise the static row is not considered present:

> INSERT INTO t (p, s) VALUES (1,NULL) IF NOT EXISTS;
+-------------+-----+------+------+------+
| [applied]   |   p | c    | s    | r    |
|-------------+-----+------+------+------|
| True        |   1 | null | null | null |
+-------------+-----+------+------+------+
> INSERT INTO t (p, s) VALUES (1,NULL) IF NOT EXISTS;
+-------------+-----+------+------+------+
| [applied]   |   p | c    | s    | r    |
|-------------+-----+------+------+------|
| True        |   1 | null | null | null |
+-------------+-----+------+------+------+

It is OK to us a comparison with NULL in a condition. But since NULL value and missing value in ScyllaDB are indistinguishable, conditions which compare with NULL will return the same result when applied to both missing rows or existing rows with NULL cells:

> UPDATE t SET s=2 WHERE p=1 IF s = NULL;
+-------------+------+
| [applied]   | s    |
|-------------+------|
| True        | null |
+-------------+------+

If a regular row is missing, but the static row cells are assigned, the static cells will be present in the row used to evaluate the condition of the “missing” regular row:

> UPDATE t SET r=2 WHERE p=1 AND c=2 IF s = 2;
+-------------+-----+
| [applied]   |   s |
|-------------+-----|
| True        |   2 |
+-------------+-----+

ScyllaDB Paxos¶

The statements with an IF clause use a different write path, employing the Paxos consensus algorithm (see figure) to ensure linearizability of the execution history.

In Paxos, in order to persist a change a coordinator first must create a unique time-based identifier, called a ballot, and send it to replicas. Upon receiving a ballot, replicas respond with a promise to accept a change associated with it.

A replica refuses to promise a ballot if it has already promised a newer one – this locks out concurrent modification attempts and allows the coordinator to proceed with reading and updating a row without interference. The state of the protocol is persisted in system.paxos table, which is local to each replica.

Unlike Cassandra, ScyllaDB piggy-backs the old version of the row on response to “Prepare” request, so reading a row doesn’t require a separate message exchange.

Once the coordinator gets a majority of promises from replicas, it evaluates the IF conditions, and if the result is true, sends an updated mutation to replicas. Replicas store the new row in system.paxos and acknowledge accepting it.

Having a majority of replicas accept the row satisfies the “quorum intersection” rule: as long as at least a majority of nodes are up, each conceivable new quorum contains at least one node which accepted the previous row and is thus aware of the change.

After the coordinator has received the majority of accepts, Paxos protocol round is complete, and it is safe to update the base table with the new row. This is done in “Learn” round.

If the base table update is successful, the coordinator responds to the client. It’s also safe to prune the state of the protocol from system.paxos.

The size of the quorum impacts how many acknowledgements the coordinator must get before proceeding to the next round or responding to the client. For Prepare and Accept, it is configured with SERIAL CONSISTENCY setting. For Learn, ScyllaDB’s eventual CONSISTENCY is used. Pruning is done in the background.

Key differences between ScyllaDB and Cassandra Paxos implementations are in collapsing prepare and read actions into a single round, and also introducing an extra asynchronous “prune” round, which keeps system.paxos table small and thus reduces write amplification when it’s compacted.

Batch statements¶

BATCH statements may contain one or more conditional statements and as such, these batches are called conditional batches. The entire conditional batch has an isolated view of the database and is executed using all-or-nothing principle. In many ways, conditional batches are similar to ACID transactions in relational databases, with the exception that a batch is executed only if all conditions in all statements are true, if not it does nothing.

Reading with Paxos¶

For queries against a single partition, it’s possible to use Paxos for reading. Paxos reads are useful if it’s necessary to retrieve the most up to date version of a row: a simple QUORUM read may not see a value that is currently being updated by a conditional write. In order to ensure a read is serial, set CONSISTENCY to SERIAL in cqlsh or prepared statement properties.

SELECT * FROM employees WHERE firstname = 'John' AND lastname = 'Doe'

CQL examples with lightweight transactions¶

How is IF different from WHERE?¶

One may think that IF clause should be used in place of WHERE - and this is true to a large extent, both accept expressions and are applied to the searched row.

Unlike the WHERE clause, IF conditions never use a secondary index; as the rows are fetched before a condition is evaluated.

Keep in mind that the IF condition applies only to a fully qualified row, meaning you must specify the partition key and in many cases the clustering key as well. In the WHERE clause you would use DELETE or UPDATE or in SET. When using UPDATE and INSERT with WHERE you would also need the VALUES clause.

If your data selection fetches multiple rows, your IF condition cannot be ambiguous, meaning it can not evaluate to TRUE for one row and to FALSE for another. This includes statements which restrict only the partition key, and not the clustering key, or the partition key and multiple clustering keys (pk = ? and ck IN (?, ?, ?). In these instances, only the conditions on static cells are accepted.

Working with conditional statement result set¶

A conditional statement always returns a result set. The result set’s first column [applied] is special: it indicates whether the statement was applied or not. Following the [applied] column are the primary key columns and then the columns used in the conditional expression. The result set contains previous (old) values for these columns. If condition is IF EXISTS or IF NOT EXISTS all columns of the column family become part of the result set. If the statement inserts a new row so previous values did not exist, the result will contain NULLs.

A batch statement may contain multiple conditional statements. Each conditional statement of a batch yields a row of the result set, and these rows are returned in statement order. Non-conditional statements may be present in a conditional batch, but do not contribute to the result set. The set of columns of conditional batch result set is a set union of all columns used in conditional expressions of all conditional statements of the batch.

BEGIN BATCH
    UPDATE movies.nowshowing SET main_actor = 'Aldis Hodge' WHERE movie = 'Invisible Man' IF main_actor = 'Elisabeth Moss'
    UPDATE movies.nowshowing SET director = 'Mr Saw ' WHERE movie = 'Invisible Man' IF director = 'Leigh Whannell'
APPLY BATCH;

 [applied] | movie         | location | run_day | run_time | director       | main_actor
-----------+---------------+----------+---------+----------+----------------+----------------
      True | Invisible Man |     null |    null |     null | Leigh Whannell | Elisabeth Moss
      True | Invisible Man |     null |    null |     null | Leigh Whannell | Elisabeth Moss

Error handling¶

Error executing a conditional statement (LWT) does not necessarily mean it failed. In addition to standard causes for failure such as client-side timeouts, server overload, memory allocation error, Paxos has its own scenarios leading to uncertainty:

• An update started by one node is allowed to be overtaken and completed by another node.

• The coordinator may fail or timeout after a write has succeeded at the majority of the participants and thus has de-facto committed.

To know for sure, a client must read the value back or retry the operation until it succeeds. Retrying should be done only if the failure is transient. It would be useless to retry in case of, e.g., a syntax error. Let’s consider some transient errors in more detail.

Limitations¶

The following limitations apply:

• You cannot use conditional batches to modify multiple partitions

• You cannot use LWT with tables which use counters

• When using user-supplied timestamps, you must make sure that the timestamp is assigned by the transaction coordinator. If not, it will not be possible to guarantee consistency.

• It is not recommended to use conditional (LWT) and non-conditional statements with the same data set as it is not possible to ensure that conditional statements are consistent in this case. In short, all LWT data sets should not use write operations that are not LWT in structure.

Other limitations are more minor:

• While a non-LWT batch can be UNLOGGED, a conditional batch cannot;

• IF conditions must be a perfect conjunct (… AND … AND …);

• Conditional batches are always logged in system.paxos table, so UNLOGGED keyword is silently ignored for them.

Additional Information¶

• How does ScyllaDB LWT Differ from Apache Cassandra ? - How does ScyllaDB’s implementation of lightweight transactions differ from Apache Cassandra?

• How to Change gc_grace_seconds for a Table - How to change the gc_grace_seconds parameter for a table