Temporal Data

Modeling when facts are true (valid-time), when they were recorded (transaction-time), or both (bitemporal), enabling time-travel queries and regulatory audit.

When to Use

• Insurance policies with effective dates that change over time
• Price histories where "what was the price on date X?" is a query requirement
• Employee role and salary changes tracked for HR reporting
• Regulatory requirements to reconstruct past states (financial compliance, audit)
• Any domain where historical accuracy matters and overwrites destroy information

Instructions

Key Concepts

Three temporal dimensions:

1. Valid-time (application time): When the fact is true in the real world. An employee salary might be effective 2024-01-01 to 2024-12-31. The database might record this fact on 2023-12-15 -- the valid time and the recording time differ.

2. Transaction-time (system time): When the row was stored in the database. Columns like

recorded_at

and

superseded_at

track the database's knowledge. Transaction-time columns are never manually edited -- the system manages them.

3. Bitemporal: Both dimensions combined. Answers questions like "what did the system believe was true about time T, as of database time S?" Essential for late-arriving corrections in financial systems.

Slowly Changing Dimensions (SCD) types:

Type	Strategy	History?
Type 1	Overwrite the row	No
Type 2	New row with date range	Yes
Type 3	Previous/current columns	Partial
Type 6	Hybrid (Type 1 + 2 + 3)	Yes

SCD Type 2 is the most common for temporal data. In PostgreSQL, use range types:

CREATE EXTENSION IF NOT EXISTS btree_gist;

CREATE TABLE employee_salaries (
  id          serial PRIMARY KEY,
  employee_id int NOT NULL,
  salary      numeric(10,2) NOT NULL,
  valid_range tstzrange NOT NULL,
  EXCLUDE USING gist (
    employee_id WITH =,
    valid_range WITH &&
  )
);

The exclusion constraint prevents overlapping valid periods for the same employee -- the database enforces temporal integrity, not the application.

Worked Example

Insurance policy pricing with valid-time tracking:

-- Initial price: $500/month effective Jan 1 to Dec 31
INSERT INTO policy_prices (policy_id, monthly_rate, valid_range)
VALUES (101, 500.00, '[2024-01-01, 2025-01-01)');

-- Price change mid-year: close the old range, insert new
UPDATE policy_prices
SET valid_range = '[2024-01-01, 2024-07-01)'
WHERE policy_id = 101 AND valid_range @> '2024-06-15'::timestamptz;

INSERT INTO policy_prices (policy_id, monthly_rate, valid_range)
VALUES (101, 550.00, '[2024-07-01, 2025-01-01)');

Point-in-time query -- "what was the price on June 15?":

SELECT monthly_rate FROM policy_prices
WHERE policy_id = 101
  AND valid_range @> '2024-06-15'::timestamptz;
-- Returns: 500.00

Bitemporal query -- "what did we think the price was on June 15, before the July correction?":

SELECT monthly_rate FROM policy_prices
WHERE policy_id = 101
  AND valid_range @> '2024-06-15'::timestamptz
  AND recorded_at < '2024-07-01'::timestamptz;

This is essential when auditors need to reconstruct what the system reported at a specific point in database history.

Anti-Patterns

1. Using

   updated_at

   as a substitute for temporal modeling.

   updated_at

   only records the last change -- all history is lost. It cannot answer "what was true at time T?"
2. Overlapping valid periods without exclusion constraints. Two rows claiming the same employee had different salaries on the same date is corrupt data. Use exclusion constraints.
3. Managing two timestamp columns manually instead of range types. Separate

   valid_from

   and

   valid_to

   columns require application-level overlap checks. Range types with exclusion constraints push this to the database.
4. Using SCD Type 1 (overwrite) when audit trail is required. Overwriting destroys evidence. Use Type 2 or a separate audit table.
5. NULL for "current" end date. Use range type upper-bound infinity instead:

   '[2024-07-01,)'

   -- the open upper bound means "still current" and works correctly with range operators.

PostgreSQL Specifics

Range types (

tstzrange

,

daterange

) provide built-in temporal operators:

• @>

  -- contains a timestamp:

  valid_range @> '2024-06-15'::timestamptz

• &&

  -- overlaps another range:

  range1 && range2

• -|-

  -- adjacent ranges:

  range1 -|- range2

GiST index for range columns:

CREATE INDEX idx_prices_valid ON policy_prices USING gist (valid_range);

Exclusion constraints require the

btree_gist

extension:

CREATE EXTENSION btree_gist;

-- Prevents overlapping valid periods for the same policy
EXCLUDE USING gist (policy_id WITH =, valid_range WITH &&)

SQL:2011 temporal support is partially implemented in PostgreSQL. System-versioned tables are not yet native -- use triggers or the

temporal_tables

extension to approximate transaction-time tracking.

Details

Advanced Topics

SQL:2011 standard temporal syntax (

FOR SYSTEM_TIME AS OF

) is supported in MariaDB and SQL Server but not yet in PostgreSQL. PostgreSQL approximates this with view-based patterns or the

temporal_tables

extension.

Gap detection in temporal data -- finding periods with no valid record:

SELECT upper(a.valid_range) AS gap_start, lower(b.valid_range) AS gap_end
FROM policy_prices a
JOIN policy_prices b ON a.policy_id = b.policy_id
  AND upper(a.valid_range) < lower(b.valid_range)
WHERE NOT EXISTS (
  SELECT 1 FROM policy_prices c
  WHERE c.policy_id = a.policy_id
    AND c.valid_range && tstzrange(upper(a.valid_range), lower(b.valid_range))
);

Temporal joins for combining two temporal tables:

SELECT e.name, s.salary, d.department_name
FROM employees e
JOIN employee_salaries s ON e.id = s.employee_id
JOIN employee_departments d ON e.id = d.employee_id
  AND s.valid_range && d.valid_range;

Partitioning temporal tables by

valid_from

enables archival of old time periods and efficient pruning of historical queries.

Engine Differences

MySQL lacks range types and exclusion constraints entirely. Temporal data in MySQL requires:

• Separate

  valid_from DATETIME

  and

  valid_to DATETIME

  columns
• Application-level overlap prevention (no database-enforced exclusion)
• Manual NULL handling for "current" records (no range infinity)

MySQL 8.0 does not support SQL:2011 temporal tables.

MariaDB 10.3+ is the notable exception -- it supports system-versioned tables (

WITH SYSTEM VERSIONING

), making it the only major open-source engine with native transaction-time temporal support. MariaDB's temporal tables automatically track row history without triggers.

Real-World Case Studies

Financial compliance system tracking 10M account balances. Regulators require reconstructing the account balance at any historical date. SCD Type 2 implementation using

tstzrange

with GiST index on

(account_id, valid_range)

. Point-in-time queries ("what was balance on date X?") complete in 3ms. After an audit discovered late-arriving corrections (a deposit backdated to last month), the team added bitemporal columns to track both valid-time and transaction-time. This enabled the query "what did we report the balance was on March 15, before the April 2 correction?" -- critical for regulatory reconciliation.

Source

• PostgreSQL Range Types
• Snodgrass -- Developing Time-Oriented Database Applications

Process

1. Read the key concepts to understand the three temporal dimensions and SCD types.
2. Apply the correct temporal strategy based on whether you need valid-time, transaction-time, or bitemporal tracking.
3. Verify that range types with exclusion constraints prevent overlapping periods and that point-in-time queries return correct results.

Harness Integration

• Type: knowledge -- this skill is a reference document, not a procedural workflow.
• No tools or state -- consumed as context by other skills and agents.
• related_skills: db-audit-trail, db-time-series, db-denormalization, microservices-event-sourcing

Success Criteria

• Temporal tables use range types with exclusion constraints (not manual timestamp pairs).
• SCD type is chosen based on business requirements (Type 2 for full history, Type 1 only when history is not needed).
• Point-in-time queries are verified to return correct results.
• GiST indexes are present on range columns.