Chronos/DBWaitlist

Chronos DB, self-hosted

World sets and objectives

Fork many worlds from one base, read and compare them, rank them, merge the best into the base; objectives and OPTIMIZE. Next release.

Self-hostedChronos DB is one binary you run yourself. Releases and install script. To put agents on a database you already run, with approvals: Chronos Gateway

Which release

What a world set is

A world set is a base world and worlds forked from it: many tries at one task, kept side by side, compared, ranked, and the best merged back into the base. The base can’t be main: the best world merges into the base, and the base merges into main as any world does, under its rules.

a whole round
create world plan;create world set pricing from world plan with (max_live = 100);fork world set pricing count 50 run $$update prices set p = p * (0.9 + random() * 0.2) where sku % 50 = $1$$ seed 42;adopt world my_try into set pricing;                   -- a fork of the base, made some other wayselect _world, sum(p * sold) from prices at worlds of set pricing group by 1;compare worlds pricing_3, pricing_7;                   -- table, key, a_value, b_valuerank world set pricing by ($$select sum(p * sold) from prices$$ desc, $$select count(*) from stockouts$$ asc)  with checks;                                         -- world, rank, score, scores, failed, reason, keptprune world set pricing where rank > 5 or failed;      -- the worlds discardedmerge best of set pricing into plan dry run;           -- world, action, rowsmerge best of set pricing into plan;

Make a set

  • CREATE WORLD SET s FROM WORLD b [WITH (max_live = n)] names the set and its base. max_live limits the set’s worlds (53400 past it).
  • FORK WORLD SET s COUNT n RUN $$script$$ [SEED k] [THREADS t] forks n worlds s_<index> and runs the script in each, as SIMULATE does: $1 is the index, $2 the seed, random() is seeded, now() is the fork’s moment. It keeps all of them, failed ones too.
  • The same every time. The same seed gives the same worlds on any number of threads, and REPLAY WORLD makes one again.
  • ADOPT WORLD w INTO SET s adds a fork of the base made some other way: by hand, by an agent.

Read every world of a set

AT WORLDS OF SET s after a table reads it in every world of the set, one after another, each row with a last column _world naming its world.

  • Same columns: the worlds must give the table the same columns. 42804 names the world that does not: read it alone with AT WORLD.
  • An empty set gives the base’s columns and no rows.
  • Read only, with the reader’s rights, tenant and masks in each world, as AT WORLD.

COMPARE WORLDS

COMPARE WORLDS
compare worlds pricing_3, pricing_7;           -- table, key, a_value, b_valuecompare worlds pricing_3, pricing_7 summary;   -- table, a_changed, b_changed, differing

The rows the two worlds hold differently. Two worlds of one set are compared by what each changed since the fork, not row by row.

RANK and PRUNE

  • Metrics: each is one SELECT giving a number, run read only in each world, as the caller. The first metric orders the worlds, the next break ties, then the index.
  • Last place: a world whose script failed, whose metric failed, or (with WITH CHECKS, or WITH CHECKS (c, ...) for some) that breaks a merge check as a merge into the base would run it.
  • Kept: the ranking is stored in each world’s metadata (_sonos_set_scores).
  • PRUNE WORLD SET s WHERE rank > k OR failed discards by the kept ranking and returns the worlds discarded.
  • A world changed since it was ranked makes PRUNE and MERGE BEST refuse (40001): rank again. A merge into a world moves its version on as a write does, so a ranking sees it.

MERGE BEST

MERGE BEST OF SET s INTO b [DRY RUN] merges rank 1 into the base with an ordinary MERGE WORLD as the caller, under the caller’s rights, merge policy and merge checks, then discards the rest. It never merges straight into main. DRY RUN returns world, action and rows and changes nothing.

Objectives and OPTIMIZE

an objective
create objective holding_cost minimize $$select sum(qty * unit_cost) from stock$$ subject to checks (no_stockout);optimize holding_cost from world plan as set hc run $$update stock set qty = qty - $1 % 7$$ count 200 keep 5 seed 42;drop objective holding_cost;
  • An objective is what OPTIMIZE ranks by, and the merge checks a world must pass. Objectives are kept on main; only the database’s own users and admins make them.
  • OPTIMIZE makes a world set (named by AS SET, else after the objective), forks and runs COUNT worlds, ranks them by the objective with its checks, and with KEEP k discards all but the best k that passed.
  • Without SUBJECT TO CHECKS an objective judges no merge checks.
  • Then: MERGE BEST OF SET hc INTO plan, as for any set.

Agents and limits

WhatHow
Forked for an agentEach world is forked as the agent forks any world: its own, with its tenant and its merge policy’s settings. A world with no scope of its own gets the base’s.
TransactionsNone of these run in a transaction (25001).
Sizemax_live per set; the agent’s max_worlds, max_children and budgets as for any fork.
Errors53400 past max_live; 40001 for a world changed since its ranking; 42804 for a set whose worlds disagree on a table’s columns; 25001 in a transaction.

Next

World sets and objectives · Chronos