temporal-data

Temporal Data Patterns

Patterns for implementing temporal databases with system-time versioning.

Temporal Dimensions

Temporal databases track data along one or two time dimensions:

• System time (transaction time): When the database recorded the fact. Managed automatically; never user-editable. This is the focus of this skill.
• Valid time (business time): When the fact was true in reality. Application-managed; can be backdated or future-dated.
• Bitemporal: Both dimensions. Enables "what did the database think at time T1 about the state at time T2?"

This skill covers system time only. For valid time and bitemporal patterns, see Liftwizard temporal docs and the Klass DSL.

Core Principles

• Immutability: All writes into the data store are immutable and append-only, except for the system_to value.
• Contiguous timeline: system_from of a new row equals system_to of the old row. This forms an unbroken chain of versions per entity.
• Far-future sentinel over NULL: Use a far-future date (e.g. 9999-12-31 23:59:59) instead of NULL for open-ended records. Enables NOT NULL constraints, composite primary keys, and uniform query syntax.
• Deduplication: Before performing a phase-out/phase-in, compare the incoming data with the current row. If unchanged, leave the row untouched. This avoids creating adjacent rows with identical data for different time ranges, which would produce false changes in history queries and diffs.

Schema Design

Temporal Columns

Every temporal table has two additional columns:

Column	Semantics	Default
system_from	When this version became valid	Current timestamp (NOT NULL)
system_to	When this version was superseded	Far-future sentinel (NOT NULL)

Primary Key

Include a temporal column in the primary key to allow multiple versions per entity.

Recommended: PK on (id, system_to) (Reladomo convention)

The rationale is bug detection: current records all share system_to = FAR_FUTURE_DATE, so a bug that creates duplicate "current" records for the same ID triggers an immediate PK violation. With (id, system_from), duplicate-current-record bugs only surface if two records have the exact same start timestamp — unlikely, so bugs go undetected.

If both constraints are needed, use PK on one and a unique index on the other.

Indexes

Every temporal table needs an index on (id, system_to) for the most common query pattern: finding the current version via WHERE id = ? AND system_to = FAR_FUTURE_DATE.

Example DDL

CREATE TABLE nodes (
  id TEXT NOT NULL,
  name TEXT,
  system_from TEXT NOT NULL DEFAULT (strftime('%Y-%m-%d %H:%M:%S', 'now')),
  system_to TEXT NOT NULL DEFAULT '9999-12-31 23:59:59',
  PRIMARY KEY (id, system_to)
);

CREATE INDEX nodes_system_from_idx ON nodes(id, system_from);

Adding Temporal Columns to Existing Table

SQLite requires recreating the table to change the primary key:

CREATE TABLE my_table_new (
  id TEXT NOT NULL,
  -- existing columns...
  system_from TEXT NOT NULL,
  system_to TEXT NOT NULL DEFAULT '9999-12-31 23:59:59',
  PRIMARY KEY (id, system_to)
);

INSERT INTO my_table_new
SELECT *, strftime('%Y-%m-%d %H:%M:%S', 'now'), '9999-12-31 23:59:59'
FROM my_table;

DROP TABLE my_table;
ALTER TABLE my_table_new RENAME TO my_table;

CREATE INDEX my_table_system_from_idx ON my_table(id, system_from);

For databases that support ALTER TABLE ... ADD PRIMARY KEY, this is simpler:

ALTER TABLE my_table ADD COLUMN system_from TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP;
ALTER TABLE my_table ADD COLUMN system_to TIMESTAMP NOT NULL DEFAULT '9999-12-31 23:59:59';
ALTER TABLE my_table DROP PRIMARY KEY;
ALTER TABLE my_table ADD PRIMARY KEY (id, system_to);
CREATE INDEX my_table_system_from_idx ON my_table(id, system_from);

After migration, update application code to add WHERE system_to = '9999-12-31 23:59:59' to all existing queries.

Query Patterns

Current Records

SELECT * FROM nodes WHERE system_to = '9999-12-31 23:59:59';

-- For a specific entity
SELECT * FROM nodes WHERE id = ? AND system_to = '9999-12-31 23:59:59';

As-Of Query (Point-in-Time)

Use strict inequality: system_from <= T AND system_to > T

SELECT * FROM nodes
WHERE id = ?
  AND system_from <= '2024-06-15 14:30:00'
  AND system_to > '2024-06-15 14:30:00';

This ensures exactly one row matches at any point in time. When rows have contiguous timestamps (system_to of row N = system_from of row N+1), <= on both sides (BETWEEN) would match two rows at the boundary. The strict > on system_to prevents this.

WRONG: Do not use BETWEEN for as-of queries.

-- WRONG: matches two rows at boundary timestamps
WHERE system_from BETWEEN '2024-01-01' AND '2024-12-31'

History of an Entity

SELECT * FROM nodes WHERE id = ? ORDER BY system_from DESC;

Entities Changed in a Range

SELECT DISTINCT id FROM nodes
WHERE system_from >= '2024-01-01' AND system_from <= '2024-12-31';

Write Operations

Phase Out and Replace (Non-Destructive Update)

-- Step 1: Close the current version
UPDATE nodes
SET system_to = '2024-06-15 14:30:00'
WHERE id = 'node-1' AND system_to = '9999-12-31 23:59:59';

-- Step 2: Insert the new version
INSERT INTO nodes (id, name, system_from, system_to)
VALUES ('node-1', 'New Name', '2024-06-15 14:30:00', '9999-12-31 23:59:59');

The new row's system_from must equal the old row's system_to to maintain a contiguous timeline. Both statements must run in the same transaction.

Non-Destructive Delete

Phase out without inserting a replacement:

UPDATE nodes
SET system_to = '2024-06-15 14:30:00'
WHERE id = 'node-1' AND system_to = '9999-12-31 23:59:59';

Merge List Pattern (Three-Way Sync)

When syncing from an external source to a temporal cache, implement ALL THREE legs:

External Source         Cache (temporal)
+-------------+         +-------------+
| A (updated) |         | A (old)     |  <- LEG 1: Update
| B (same)    |         | B (same)    |  <- LEG 1: Leave untouched
| C (new)     |         | D (deleted) |  <- LEG 2: Insert C
+-------------+         +-------------+  <- LEG 3: Phase out D

Leg 1: Sync Existing (matching IDs in both source and cache)

Compare data. If changed, phase out old version and insert new. If unchanged, leave untouched (do nothing).

Leg 2: Add New (in source but not in cache)

Insert new records with system_from = now, system_to = FAR_FUTURE_DATE.

Leg 3: Phase Out Orphaned (in cache but not in source) — COMMONLY FORGOTTEN

UPDATE nodes
SET system_to = @now
WHERE system_to = '9999-12-31 23:59:59'
  AND id NOT IN (SELECT id FROM source_ids);

Common Bug: Missing Leg 3

The most common temporal sync bug is forgetting Leg 3. Symptoms:

• Deleted items remain in cache forever
• Cache grows unbounded
• Stale data causes errors when writing back to source

Sync Checklist

Before considering a sync function complete:

• Leg 1: Updates existing records when data changes
• Leg 2: Inserts records not in cache
• Leg 3: Phases out cached records not in source
• Transaction wraps all three legs atomically
• Handles empty source list (phases out all cached records)

Bulk Import Pattern

For bulk imports from an external source, apply the same compare-before-cut logic as the Merge List Pattern:

For each imported record, compare it against the current row in the temporal table:

• Data changed — phase out the old row and insert a new version
• Data unchanged — do nothing (leave the existing row untouched)
• New record (no current row) — insert it
• Missing from import (current row has no match in the import) — phase out

This avoids creating unnecessary versions for unchanged data and provides clear categorization: added / updated / unchanged / deleted.

Multi-Table / Parent-Child

Each temporal table must independently track its own system_from/system_to columns. You cannot derive a child table's system_to by joining on the parent's system_from, because deduplication means the parent row may not change when only child data changes (and vice versa).

When several tables are edited within a single transaction, all system_from/system_to values use the same timestamp. This keeps the temporal timeline consistent across related tables and allows as-of queries at that timestamp to return a coherent snapshot.

Unchanged data is copied from the old row to the new row. For very wide columns that don't change frequently, it may be more efficient to split them out into a separate table. This reduces row duplication overhead when only a small part of the entity changes.

Temporal Rollback (Disaster Recovery)

To roll back the database to a previous point in time, apply two operations per table:

-- 1. Purge: Delete rows created after the rollback point
DELETE FROM nodes WHERE system_from > @target_timestamp;

-- 2. Restore: Re-open rows that were current at the rollback point
UPDATE nodes
SET system_to = '9999-12-31 23:59:59'
WHERE system_from <= @target_timestamp
  AND system_to > @target_timestamp;

This is a destructive operation that violates the immutability principle. Use only for disaster recovery.

Composites

A composite is an entity composed of multiple tables. For example, a Blueprint composite might include Blueprint, BlueprintTag, and ImgurImage tables.

Ownership direction matters. BlueprintTag is part of the Blueprint composite, but it is NOT part of the Tag composite. A tag is a standalone entity that exists independently; BlueprintTag is a join table owned by Blueprint.

When any part of a composite is edited, the version number for the entire composite bumps once. A change to BlueprintTag bumps the Blueprint version, not the Tag version. The version number applies to the composite as a whole, not to individual tables within it.

This ownership model determines which version table gets updated during writes and which set of tables participate in as-of-by-version queries.

Versioning

A separate version table tracks version numbers with their own system_from/system_to:

CREATE TABLE question_version (
  question_id BIGINT NOT NULL,
  number INTEGER NOT NULL,
  system_from TIMESTAMP NOT NULL,
  system_to TIMESTAMP NOT NULL DEFAULT '9999-12-31 23:59:59',
  PRIMARY KEY (question_id, system_to)
);

Version queries work in two steps:

1. Look up the version number in the version table to get system_from
2. Use that timestamp for as-of queries on all related tables in the composite

The version table is the entry point for as-of-by-version queries. Without it, callers would need to know the exact timestamp for a given version.

-- Step 1: Find the timestamp for version 1
SELECT system_from FROM question_version
WHERE question_id = ? AND number = 1;

-- Step 2: Use that timestamp for as-of queries on all related tables
SELECT * FROM question
WHERE id = ? AND system_from <= @version_timestamp AND system_to > @version_timestamp;

SELECT * FROM question_tag
WHERE question_id = ? AND system_from <= @version_timestamp AND system_to > @version_timestamp;

Editing any table in the composite bumps the version once. The version number applies to the entire composite, not individual tables. See the Composites section above for ownership semantics.

Auditing

Auditing adds tracking of who made each change. Requires temporal support as a prerequisite.

CREATE TABLE question (
  id BIGINT NOT NULL,
  title TEXT,
  body TEXT,
  created_by_id TEXT NOT NULL,
  created_on TIMESTAMP NOT NULL,
  last_updated_by_id TEXT NOT NULL,
  system_from TIMESTAMP NOT NULL,
  system_to TIMESTAMP NOT NULL DEFAULT '9999-12-31 23:59:59',
  PRIMARY KEY (id, system_to)
);

• created_by_id and created_on are set once on creation and never change
• last_updated_by_id is updated on each new version
• These properties should appear on both the main table and the version table

Optimistic Locking

APIs that perform edits require a version number as input and fail if the input version number doesn't match the current version. This prevents multiple users from accidentally overwriting each other's changes. When two users load the same entity, both see version N. The first user's edit succeeds and bumps the version to N+1. The second user's edit sends version N, which no longer matches, so the API rejects it with a conflict error.

See Liftwizard temporal documentation for more details.

Framework-Specific Patterns

Klass DSL

The Klass DSL uses classifier modifiers as compiler macros:

class Question
    systemTemporal
    versioned
    audited
{
    id   : Long key;
    title: String;
    body : String;
}

systemTemporal infers system/systemFrom/systemTo properties. versioned infers a version class. audited infers createdById/createdOn/lastUpdatedById properties.

Reladomo (Liftwizard)

Liftwizard uses Reladomo for temporal support. Reladomo manages system_from/system_to automatically through AsOfAttribute declarations in XML object definitions. The ORM handles phase-out/phase-in within transactions, contiguous timelines, and as-of query generation.

See Liftwizard temporal documentation for non-destructive updates, as-of queries, versioning, auditing, optimistic locking, diffs, and maker/checker workflows.

Drizzle ORM (TypeScript/SQLite)

import {sqliteTable, text, primaryKey, uniqueIndex} from 'drizzle-orm/sqlite-core';

export const nodes = sqliteTable(
	'nodes',
	{
		id: text('id').notNull(),
		name: text('name'),
		systemFrom: text('system_from').notNull(),
		systemTo: text('system_to').notNull().default('9999-12-31 23:59:59'),
	},
	(table) => ({
		pk: primaryKey({columns: [table.id, table.systemTo]}),
		systemFromIdx: uniqueIndex('nodes_system_from_idx').on(table.id, table.systemFrom),
	}),
);

Best Practices

• Always use transactions for temporal operations involving multiple statements
• Compare data before inserting to avoid duplicate versions when data hasn't changed
• Index (id, system_to) for current-record queries — this is the most important index
• Monitor database size — temporal tables grow with every change; deduplicate unchanged data to minimize bloat
• Test as-of queries to ensure correct temporal semantics — off-by-one at boundaries is a common bug
• Use prepared statements for better performance in bulk operations