Imported from ctoth/sql-skills-plugin (plugins/sql/skills/sql-match-recognize/SKILL.md). Install upstream with npx skills add ctoth/sql-skills-plugin --skill sql-match-recognize. Copyright stays with the author.

SQL MATCH_RECOGNIZE — Row Pattern Recognition

"From the SQL viewpoint, you can think of row pattern matching as extended window functions. ... Before the introduction of MATCH_RECOGNIZE, you had to feed your data to external tools to reason about trends and patterns. Now, you can achieve it directly in your query." — Trino — Row pattern matching

"[MATCH_RECOGNIZE] uses simple regular expressions across multiple rows to filter them. ... useful to implement [f]inding series of consecutive events [and p]attern matching: trend reversal, periodic events." — modern-sql.com — MATCH_RECOGNIZE

MATCH_RECOGNIZE runs a regular expression over an ordered stream of rows. You partition and sort the rows (exactly like a window), define named row conditions (DEFINE), describe the shape you want as a regex over those names (PATTERN), and project results from each match (MEASURES). Problems that otherwise demand a tangle of correlated self-joins or nested LAG/CASE — "a price dip then recovery," "login then 2+ failures then success," "three rising bars then a reversal" — collapse to a few declarative lines. This skill assumes the set semantics and three-valued logic of the foundation sql-relational-and-null-discipline (the match runs over an explicit ORDER BY; nothing is ordered without it) and the window mechanics of sql-window-functions. It is the escalation target named by sql-gaps-and-islands §6 when window tricks get unwieldy.

Read this first — the portability reality (full detail in §6): row pattern matching is an optional SQL:2016 feature. It is established on Oracle, Trino, Snowflake, Flink, Vertica, and DB2, but it is absent from the PostgreSQL, MySQL, MariaDB, and SQLite versions most people actually run today. If your engine does not have it, do not reach for it — use the window-function recipes in sql-gaps-and-islands instead.

1. What MATCH_RECOGNIZE Is — and the Fixed Clause Order

The clause sits in the FROM, wrapping a table. The sub-clauses must appear in this exact order, and only PATTERN and DEFINE are mandatory (Trino):

SELECT ...
FROM some_table
  MATCH_RECOGNIZE (
    [ PARTITION BY col [, ...] ]      -- split into independent streams (like a window partition)
    [ ORDER BY    col [, ...] ]       -- REQUIRED ordering within each partition
    [ MEASURES    expr AS name [, ...] ]  -- columns computed from a match
    [ ONE ROW PER MATCH | ALL ROWS PER MATCH ]  -- output granularity (default ONE ROW)
    [ AFTER MATCH SKIP ... ]          -- where to resume after a match
    PATTERN ( row_pattern )           -- the regex over pattern variables  (MANDATORY)
    [ SUBSET ... ]                    -- union of variables, for MEASURES navigation
    DEFINE  var AS condition [, ...]  -- the boolean test each variable's row must pass (MANDATORY)
  );

Map it to a regex you already know:

• DEFINE is the alphabet: each pattern variable is a boolean predicate a row must satisfy to "be" that symbol (DOWN AS price < PREV(price)). A variable used in PATTERN but never defined matches any row (a useful anchor).
• PATTERN is the regular expression over those symbols: START DOWN+ UP+ means "one anchor row, then one-or-more DOWN rows, then one-or-more UP rows."
• MEASURES projects the result: navigate within the match with FIRST/LAST/PREV/NEXT and var.column (LAST(DOWN.price) = the price at the last DOWN row).
• PARTITION BY/ORDER BY structure the input "similarly to window functions" (Trino).

PREV(price) reads the prior row's value — pattern variables are evaluated as the match is being built (running semantics, §4).

2. The Worked Example — Detect a V-Shape (Dip and Recovery)

This is the centerpiece. Goal: per customer, find a run where the price falls for one or more steps, then rises for one or more steps — a V. First, the way it gets written without MATCH_RECOGNIZE.

BEFORE — window/self-join gymnastics

A genuine "1+ down steps then 1+ up steps, of unbounded length" cannot be expressed by a fixed LAG(price, k) chain — the dip length is variable. The portable approximations are ugly: either gaps-and-islands grouping over a per-row down/up direction flag with a running-sum group key and extra logic to require a down-run immediately followed by an up-run, or a correlated self-join that, for every candidate bottom row, scans backward for a monotone fall and forward for a monotone rise:

-- BEFORE (sketch) — find each "bottom" b, then prove a monotone fall into it and a monotone rise out.
-- Each EXISTS/NOT EXISTS hides a correlated scan; pinning the exact start/end of the V needs more.
-- Real-world versions of this run 60-100 lines and are a magnet for off-by-one and "what about ties" bugs.
SELECT b.customer_id, b.order_date AS bottom_date, b.price AS bottom_price
FROM   orders b
WHERE  EXISTS (SELECT 1 FROM orders d           -- a strictly-lower predecessor exists
               WHERE d.customer_id = b.customer_id AND d.order_date < b.order_date
                 AND d.price > b.price
                 AND NOT EXISTS (SELECT 1 FROM orders m  -- ...with nothing rising in between
                                 WHERE m.customer_id = b.customer_id
                                   AND m.order_date BETWEEN d.order_date AND b.order_date
                                   AND m.price > /* prior row's price */ 0))   -- gets hairy fast
  AND  EXISTS (SELECT 1 FROM orders u WHERE u.customer_id = b.customer_id
               AND u.order_date > b.order_date AND u.price > b.price);
-- ...and you still have not produced start_price, final_price, or the V's start/end dates cleanly.

AFTER — MATCH_RECOGNIZE

The whole thing is the regex START DOWN+ UP+ over two row predicates (Trino):

SELECT customer_id, start_price, bottom_price, final_price, start_date, final_date
FROM orders
  MATCH_RECOGNIZE (
    PARTITION BY customer_id
    ORDER BY order_date
    MEASURES
      START.price          AS start_price,
      LAST(DOWN.price)     AS bottom_price,   -- price at the deepest matched DOWN row
      LAST(UP.price)       AS final_price,
      START.order_date     AS start_date,
      LAST(UP.order_date)  AS final_date
    ONE ROW PER MATCH
    AFTER MATCH SKIP PAST LAST ROW
    PATTERN (START DOWN+ UP+)
    DEFINE
      DOWN AS price < PREV(price),           -- this row is lower than the one before it
      UP   AS price > PREV(price)            -- this row is higher than the one before it
  );

For data where cust_1 goes 100, 200, 100, 50, 100 and cust_2 goes 8, 4, 6, this returns one row per detected V: cust_1's 200 -> 50 -> 100 and cust_2's 8 -> 4 -> 6 (Trino). START is undefined, so it matches the anchor row; DOWN+ consumes the fall; UP+ consumes the recovery. Six lines, one pass, no self-join, and the MEASURES hand you the start/bottom/final prices directly. This is the payoff in the pull-quote.

3. PATTERN Quantifiers — Greedy vs Reluctant

PATTERN is a regex over pattern variables, with the quantifiers you expect (Trino):

Concatenation (A B), alternation (A | B), grouping (( A B )), and the empty pattern (()) all work as in ordinary regex.

Quantifiers are greedy by default: "they prefer higher number of repetitions over lower number. If you want it the other way, you can change a quantifier to reluctant by appending ? immediately after it" (Trino). So DOWN+ (greedy) extends the dip as far as it can; DOWN+? (reluctant) stops at the minimum. The same ? appended to the optional quantifier flips its preference: "(pattern)? prefers a single match of the pattern, while (pattern)?? would rather omit the pattern altogether" (Trino).

-- Greedy: B+ swallows every consecutive B it can. The match is as LONG as possible.
PATTERN (A B+ C)
-- Reluctant: B+? takes the FEWEST B rows that still let C match. The match is as SHORT as possible.
PATTERN (A B+? C)

The trap: the trailing ? is overloaded. +? is "one-or-more, reluctant," not "one-or-more, then optional." Reach for reluctant when a greedy quantifier would swallow rows that belong to the next match (see common-mistakes §5).

4. MEASURES and RUNNING vs FINAL

Two evaluation times exist, and conflating them is the subtlest bug here (Trino):

• DEFINE conditions are evaluated while the match is in progress — "running" semantics. They have to be: a variable's predicate decides whether to extend the match, so it can only see rows matched so far. LAST(DOWN.price) inside DEFINE means "the most recent DOWN row up to the current one," not the eventual last one.
• MEASURES expressions are evaluated when the match is complete — by default these read "final" values for the whole match.

When you ask for one summary row (ONE ROW PER MATCH, §5), the distinction is mostly invisible — you get the completed match's values. It becomes load-bearing with ALL ROWS PER MATCH, where each matched row emits and you can choose per measure:

MEASURES
  RUNNING LAST(DOWN.price) AS running_low,   -- the lowest-so-far at THIS output row
  FINAL   LAST(DOWN.price) AS final_low      -- the match's eventual bottom, on every output row
ALL ROWS PER MATCH

RUNNING is the default in both clauses; FINAL applies only in MEASURES (Trino). Rule of thumb: in DEFINE you are always running (you cannot see the future); in MEASURES with ALL ROWS PER MATCH, say FINAL explicitly when you want the completed-match value on every row, or you will silently get the running one.

5. ONE ROW vs ALL ROWS PER MATCH, and AFTER MATCH SKIP

Output granularity. ONE ROW PER MATCH (the default) emits a single summary row per match — use it for "give me each detected V with its start/bottom/end." ALL ROWS PER MATCH emits one row per matched input row, carrying the running measures — use it when you need to see, and label, every row inside the match (e.g. tag each row with its pattern variable and the running low).

-- ONE ROW PER MATCH  -> 1 row per V         (summary: start_price, bottom_price, final_price)
-- ALL ROWS PER MATCH -> N rows per V         (every order row in the V, plus running measures)

Where matching resumes — AFTER MATCH SKIP. After a match completes, the engine restarts the search somewhere. "The default option is AFTER MATCH SKIP PAST LAST ROW, but you can also skip to the next row or to a specific position in the match based on the matched pattern variables" (Trino).

Choose deliberately: the default finds each V once; SKIP TO NEXT ROW finds every overlapping V. Warning: a SKIP TO that lands on the match's own first row is an error in conforming engines — it would loop forever — so do not skip to the anchor variable.

6. The Low-Portability Reality — and the Portable Fallback

This is the part that decides whether you can use the clause at all. Row pattern matching is an optional SQL:2016 feature (feature R010) (modern-sql.com) — "optional" means a conforming engine is not required to implement it, and most did not for years.

PORTABILITY — confirm engine support before recommending MATCH_RECOGNIZE.

Engine	MATCH_RECOGNIZE?
Oracle (12c R1+), Trino, Snowflake, Flink SQL, Vertica, DB2 (LUW 12.1.4+)	yes — established
PostgreSQL	only 18+ (absent from deployed 16/17)
MySQL	only 9.7+ (absent from 8.x)
MariaDB	only 12.3+
SQLite	only 3.53+
SQL Server	only 2025+

Versions per modern-sql.com — MATCH_RECOGNIZE. The takeaway: the analytics engines (Oracle, Trino, Snowflake, Flink, Vertica, DB2) have had it for years; the popular OLTP engines got it only in their newest, mostly-not-yet-deployed releases. Assume your Postgres/MySQL/MariaDB/SQLite does not have it unless you have checked the version.

When the engine lacks it, use the portable window-function fallback in sql-gaps-and-islands. "Consecutive runs" and "gap over a threshold" map cleanly onto the ROW_NUMBER-difference and running-sum-of-a-flag recipes there, which use only standard SQL:2003 window functions and run on PostgreSQL 8.4+, MySQL 8, SQLite 3.25+, everywhere. Those recipes get unwieldy precisely for the rich multi-state patterns MATCH_RECOGNIZE was built for (its own §6 routes here for that escalation) — which is the whole tension: the clean tool is not portable, and the portable tool is not always clean. The route matrix and per-engine version detail are owned by sql-standard-vs-dialect-map.

7. Who Suffers When This Is Done by Hand (or Reached For Blindly)

• The time-series engineer who wrote 80 lines of fragile correlated self-joins to detect a price dip-and-recovery — three passes, an off-by-one at every boundary row, and still no clean way to emit the start/bottom/final prices — when the engine running the warehouse (Trino/Snowflake/Oracle) supports the 6-line PATTERN (START DOWN+ UP+) in §2. The hand-rolled version is the bug they are now debugging under a deadline.
• The analyst who copied a slick MATCH_RECOGNIZE query from a blog into their PostgreSQL 16 instance and got a syntax error, because the feature landed only in PG 18 (§6). They needed to be routed to the gaps-and-islands window fallback before writing a line — confirming engine support is step zero, not a debugging afterthought.
• The next maintainer handed a greedy DOWN+ pattern that silently swallowed the first row of the next dip, so two adjacent Vs were reported as one. No error — just a metric that was quietly wrong. A reluctant DOWN+? or the right AFTER MATCH SKIP was the fix (§3, §5).

The check you ran ("does this engine have MATCH_RECOGNIZE?") and the reluctant quantifier you reached for are gifts to whoever audits the result under load.

8. Routing to Related Skills

• sql-relational-and-null-discipline — the foundation: a result is an unordered set, so the match runs over an explicit ORDER BY; three-valued logic governs the DEFINE predicates (a NULL comparison is UNKNOWN, so the row does not match that variable).
• sql-gaps-and-islands — the portable fallback (§6): ROW_NUMBER-difference and running-sum-of-a-flag window recipes for consecutive runs and threshold sessionization when your engine lacks MATCH_RECOGNIZE. It routes back here when the pattern outgrows those tricks.
• sql-window-functions — the window mechanics MATCH_RECOGNIZE extends: PARTITION BY/ORDER BY, LAG/LEAD (the portable cousins of PREV/NEXT), and frames.
• sql-standard-vs-dialect-map — the authoritative per-engine support matrix and version thresholds (§6), plus dialect quirks in pattern/measure spelling.

9. Reference Files

High-frequency MATCH_RECOGNIZE anti-patterns in LLM-generated SQL, each with wrong/right code and a primary-source citation:

references/common-mistakes.md

Source provenance for every claim in this skill:

references/sources.yaml