Recursive CTEs

PREVIEW Recursive CTEs using the WITH MUTUALLY RECURSIVE syntax is in public preview. It is under active development and may have stability or performance issues. It isn't subject to our backwards compatibility guarantees.

Recursive CTEs operate on the recursively-defined structures like trees or graphs implied from queries over your data.

Syntax

with_recursive_cte

WITH MUTUALLY RECURSIVE ( RETURN AT ERROR AT RECURSION LIMIT limit ) recursive_cte_binding , select_stmt

recursive_cte_binding

cte_ident ( col_ident col_type , ) AS ( select_stmt )
Field Use
RETURN AT RECURSION LIMIT $n An optional clause indicating that the fixpoint computation should stop after $n iterations and use the current values computed for each recursive_cte_binding in the select_stmt. This could be useful when debugging and validating the correctness of recursive queries, or when you know exactly how many iterations you want to have, regardless of reaching a fixpoint. See the Examples section for an example.
ERROR AT RECURSION LIMIT $n An optional clause indicating that the fixpoint computation should stop after $n iterations and fail the query with an error. Adding this clause with a reasonably high limit is a good safeguard against accidentally running a non-terminating dataflow in your production clusters.
recursive_cte_binding A binding that gives the SQL fragment defined under select_stmt a cte_ident alias. This alias can be used in the same binding or in all other (preceding and subsequent) bindings in the enclosing recursive CTE block. In contrast to the cte_binding definition, a recursive_cte_binding needs to explicitly state its type as a comma-separated list of (col_ident col_type) pairs.

Details

Within a recursive CTEs block, any cte_ident alias can be referenced in all recursive_cte_binding definitions that live under the same block, as well as in the final select_stmt for that block.

A WITH MUTUALLY RECURSIVE block with a general form

WITH MUTUALLY RECURSIVE
  -- A sequence of bindings, all in scope for all definitions.
  $R_1(...) AS ( $sql_cte_1 ),
  ...
  $R_n(...) AS ( $sql_cte_n )
  -- Compute the result from the final values of all bindings.
  $sql_body

is evaluated as if it was performing the following steps:

  1. Initially, bind $R_1, ..., $R_n to the empty collection.
  2. Repeat in a loop:
    1. Update $R_1 using the current values bound to $R_1, ..., $R_n in $sql_cte_1.
    2. Update $R_2 using the current values bound to $R_1, ..., $R_n in $sql_cte_2. Note that this includes the new value of $R_1 bound above.
    3. Update $R_n using the current values bound to $R_1, ..., $R_n in $sql_cte_n. Note that this includes the new values of $R_1, ..., $R_{n-1} bound above.
  3. Exit the loop when one of the following conditions is met:
    1. The values bound to all CTEs have stopped changing.
    2. The optional early exit condition to RETURN or ERROR AT ITERATION $i was set and we have reached iteration $i.

Note that Materialize’s ability to efficiently handle incremental changes to your inputs extends across loop iterations. For each iteration, Materialize performs work resulting only from the input changes for this iteration and feeds back the resulting output changes to the next iteration. When the set of changes for all bindings becomes empty, the recursive computation stops and the final select_stmt is evaluated.

WARNING!

In the absence of recursive CTEs, every SELECT query is guaranteed to compute its result or fail with an error within a finite amount of time. However, introducing recursive CTEs complicates the situation as follows:

  1. The query might not converge (and may never terminate). Non-terminating queries never return a result and can consume a lot of your cluster resources. See an example below.
  2. A small update to a few (or even one) data points in your input might cascade in big updates in your recursive computation. This most likely will manifest in spikes of the cluster resources allocated to your recursive dataflows. See an example below.

Examples

Let’s consider a very simple schema consisting of users that belong to a hierarchy of geographical areas and exchange transfers between each other. Use the SQL Shell to run the sequence of commands below.

Example schema

-- A hierarchy of geographical locations with various levels of granularity.
CREATE TABLE areas(id int not null, parent int, name text);
-- A collection of users.
CREATE TABLE users(id char(1) not null, area_id int not null, name text);
-- A collection of transfers between these users.
CREATE TABLE transfers(src_id char(1), tgt_id char(1), amount numeric, ts timestamp);

Example data

DELETE FROM areas;
DELETE FROM users;
DELETE FROM transfers;

INSERT INTO areas VALUES
  (1, 2    , 'Brooklyn'),
  (2, 3    , 'New York'),
  (3, 7    , 'United States of America'),
  (4, 5    , 'Kreuzberg'),
  (5, 6    , 'Berlin'),
  (6, 7    , 'Germany'),
  (7, null , 'Earth');
INSERT INTO users VALUES
  ('A', 1, 'Alice'),
  ('B', 4, 'Bob'),
  ('C', 2, 'Carol'),
  ('D', 5, 'Dan');
INSERT INTO transfers VALUES
  ('B', 'C', 20.0 , now()),
  ('A', 'D', 15.0 , now() + '05 seconds'),
  ('C', 'D', 25.0 , now() + '10 seconds'),
  ('A', 'B', 10.0 , now() + '15 seconds'),
  ('C', 'A', 35.0 , now() + '20 seconds');

Transitive closure

The following view will compute connected as the transitive closure of a graph where:

  • each user is a graph vertex, and
  • a graph edge between users x and y exists only if a transfer from x to y was made recently (using the rather small 10 seconds period here for the sake of illustration):
CREATE MATERIALIZED VIEW connected AS
WITH MUTUALLY RECURSIVE
  connected(src_id char(1), dst_id char(1)) AS (
    SELECT DISTINCT src_id, tgt_id FROM transfers WHERE mz_now() <= ts + interval '10s'
    UNION
    SELECT c1.src_id, c2.dst_id FROM connected c1 JOIN connected c2 ON c1.dst_id = c2.src_id
  )
SELECT src_id, dst_id FROM connected;

To see results change over time, you can SUBSCRIBE to the materialized view and then use a different SQL Shell session to insert some sample data into the base tables used in the view:

SUBSCRIBE(SELECT * FROM connected) WITH (SNAPSHOT = FALSE);

You’ll see results change as new data is inserted. When you’re done, cancel out of the SUBSCRIBE using Stop streaming.

NOTE: Depending on your base data, the number of records in the connected result might get close to the square of the number of users.

Strongly connected components

Another thing that you might be interested in is identifying maximal sub-graphs where every pair of users are connected (the so-called strongly connected components (SCCs)) of the graph defined above. This information might be useful to identify clusters of closely-tight users in your application.

Since you already have connected defined as a MATERIALIZED VIEW, you can piggy-back on that information to derive the SCCs of your graph. Two users will be in the same SCC only if they are connected in both directions. Consequently, given the connected contents, we can:

  1. Restrict connected to the subset of symmetric connections that go in both directions.
  2. Identify the scc of each users entry with the lowest dst_id of all symmetric neighbors and its own id.
CREATE MATERIALIZED VIEW strongly_connected_components AS
  WITH
    symmetric(src_id, dst_id) AS (
      SELECT src_id, dst_id FROM connected
      INTERSECT ALL
      SELECT dst_id, src_id FROM connected
    )
  SELECT u.id, least(min(c.dst_id), u.id)
  FROM users u
  LEFT JOIN symmetric c ON(u.id = c.src_id)
  GROUP BY u.id;

Again, you can insert some sample data into the base tables and observe how the materialized view contents change over time using SUBSCRIBE:

SUBSCRIBE(SELECT * FROM strongly_connected_components) WITH (SNAPSHOT = FALSE);

When you’re done, cancel out of the SUBSCRIBE using Stop streaming.

NOTE: The strongly_connected_components definition given above is not recursive, but relies on the recursive CTEs from the connected definition. If you don’t need to keep track of the connected contents for other reasons, you can use this alternative SCC definition which computes SCCs directly using repeated forward and backward label propagation.

Aggregations over a hierarchy

You might want to keep track of the aggregated net balance per area for a recent period of time. This can be achieved in three steps:

  1. Sum up the aggregated net balance for the set period for each user in an user_balances CTE.
  2. Sum up the user_balances of users directly associated with an area in a direct_balances CTE.
  3. Recursively add to direct_balances the indirect_balances sum of all child areas.

A materialized view that does the above three steps in three CTEs (of which the last one is recursive) can be defined as follows:

CREATE MATERIALIZED VIEW area_balances AS
  WITH MUTUALLY RECURSIVE
    user_balances(id char(1), balance numeric) AS (
      WITH
        credits AS (
          SELECT src_id as id, sum(amount) credit
          FROM transfers
          WHERE mz_now() <= ts + interval '10s'
          GROUP BY src_id
        ),
        debits AS (
          SELECT tgt_id as id, sum(amount) debit
          FROM transfers
          WHERE mz_now() <= ts + interval '10s'
          GROUP BY tgt_id
        )
      SELECT
        id,
        coalesce(debit, 0) - coalesce(credit, 0) as balance
      FROM
        users
        LEFT JOIN credits USING(id)
        LEFT JOIN debits USING(id)
    ),
    direct_balances(id int, parent int, balance numeric) AS (
      SELECT
        a.id as id,
        a.parent as parent,
        coalesce(sum(ub.balance), 0) as balance
      FROM
        areas a
        LEFT JOIN users u ON (a.id = u.area_id)
        LEFT JOIN user_balances ub ON (u.id = ub.id)
      GROUP BY
        a.id, a.parent
    ),
    indirect_balances(id int, parent int, balance numeric) AS (
      SELECT
        db.id as id,
        db.parent as parent,
        db.balance + coalesce(sum(ib.balance), 0) as balance
      FROM
        direct_balances db
        LEFT JOIN indirect_balances ib ON (db.id = ib.parent)
      GROUP BY
        db.id, db.parent, db.balance
    )
  SELECT
    id, balance
  FROM
    indirect_balances;

As before, you can insert the example data and observe how the materialized view contents change over time from the psql with the \watch command:

SELECT id, name, balance FROM area_balances JOIN areas USING(id) ORDER BY id;
\watch 1

Non-terminating queries

Let’s look at a slight variation of the transitive closure example from above with the following changes:

  1. All UNION operators are replaced with UNION ALL.
  2. The mz_now() < ts + $period predicate from the first UNION branch is omitted.
  3. The WITH MUTUALLY RECURSIVE clause has an optional ERROR AT RECURSION LIMIT 100.
  4. The final result in this example is ordered by src_id, dst_id.
WITH MUTUALLY RECURSIVE (ERROR AT RECURSION LIMIT 100)
  connected(src_id char(1), dst_id char(1)) AS (
    SELECT DISTINCT src_id, tgt_id FROM transfers
    UNION ALL
    SELECT src_id, dst_id FROM connected
    UNION ALL
    SELECT c1.src_id, c2.dst_id FROM connected c1 JOIN connected c2 ON c1.dst_id = c2.src_id
  )
SELECT src_id, dst_id FROM connected ORDER BY src_id, dst_id;

After inserting the example data you can observe that executing the above statement returns the following error:

ERROR:  Evaluation error: Recursive query exceeded the recursion limit 100. (Use RETURN AT RECURSION LIMIT to not error, but return the current state as the final result when reaching the limit.)

The recursive CTE connected has not converged to a fixpoint within the first 100 iterations! To see why, you can run variants of the same query where the

ERROR AT RECURSION LIMIT 100

clause is replaced by

RETURN AT RECURSION LIMIT $n -- where $n = 1, 2, 3, ...

and observe how the result changes after $n iterations.

 src_id | dst_id
--------+--------
 A      | B
 ...
 src_id | dst_id
--------+--------
 A      | B
 A      | B
 ...
 src_id | dst_id
--------+--------
 A      | B
 A      | B
 A      | B
 ...

Changing the UNION to UNION ALL in the connected definition caused a full copy of transfer to be added to the current value of connected in each iteration! Consequently, connected never stops growing and the recursive CTE computation never reaches a fixpoint.

Queries with “update locality”

The examples presented so far have the following “update locality” property:

NOTE: A change in a source collection will usually cause a bounded amount of changes to the contents of the recursive CTE bindings derived after each iteration.

For example:

  • Most of the time, inserting or removing transfers will not change the contents of the connected collection. This is true because:
    • An alternative path from x to y already existed before the insertion, or
    • An alternative path from x to y will exist after the removal.
  • Inserting or removing transfers will not change most of the contents of the area_balances collection. This is true because:
    • Areas are organized in a hierarchy with a maximum height h.
    • A single transfer contributes directly only to the area_balances of the areas where the src_id and tgt_id belong.
    • A single transfer contributes indirectly only to the area_balances of the areas ancestor areas of the above two areas.
    • Consequently, each transfer change will affect at most 2 areas in each iteration and at most 2*h areas in total.

However, note that not all iterative algorithms have this property. For example:

  • In a naive PageRank implementation, the introduction of a link between two pages x and y will change the PageRank of these two pages and every page z transitively connected to either x or y. Since most graphs exhibit small-world properties, this might represent most of the existing pages.
  • In a naive k-means clustering algorithm, inserting or removing a new data point will affect the positions of the k mean points after each iteration.

Depending on the size and update frequency of your input collections, expressing algorithms that violate the “update locality” property (such as the examples above) using recursive CTEs in Materialize might lead to excessive CPU and memory consumption in the clusters that compute these recursive queries.

Back to top ↑