Chronos/DBWaitlist

Changelog / Chronos DB

Chronos DB 0.6.2

Strict proposals for payments, safer attached databases, and fixes from an audit of 0.6.0

Beta10 changesDownloads

Shipped with Chronos Gateway 0.4.4, Chronos Gateway 0.4.3, Chronos Gateway 0.4.2, Chronos Gateway 0.4.1.

Strict proposals for payments, safer attached databases, and fixes from an audit of 0.6.0. A proposal that would send a payment, payout, transfer or refund is now strict: only a person approves it, within 15 minutes, and any commit to main since it was proposed supersedes it. Beside Postgres or MySQL, merges into main are refused when the source may lose a commit it acknowledged; beside MongoDB, merges commit to a majority, journaled, and a merge whose outcome was lost is no longer dropped when it in fact committed. chronos refresh --check compares main with the source and changes nothing. Nothing changes in how rows are stored: 0.6.2 opens and writes data format 8 as 0.6.0 does. Built and signed on the maintainer's Mac as 0.6.1 was.

  • Attached databases: merges into main are refused when the source may lose a commit it acknowledged (Postgres with fsync or synchronous_commit off; MySQL with innodb_flush_log_at_trx_commit not 1, or sync_binlog not 1 with the binary log on), nothing written, saying which setting. CHRONOS_ATTACH_UNSAFE_SOURCE=1 merges anyway.
  • Attached MongoDB: a merge's commit is acknowledged by a majority of the replica set, journaled, and its reads use read concern majority. The probe that settles a merge whose commit outcome was lost now runs in a transaction it always aborts. The old probe took a merge still committing for one that never would, and dropped it, so a merge that then committed in MongoDB never reached main; it also could leave a probe document behind. A probe that can't tell yet (the merge's transaction still open) keeps the merge in doubt, and merges into main wait.
  • chronos refresh <folder> --check: the reconciliation scan. Compares main with Postgres as a refresh does and takes nothing: per table, rows only in Postgres, rows that differ, rows only in main, and tables whose columns differ. Exit 1 if they differ. Postgres sources only for now.
  • Fixes from an audit of 0.6.0.
    • A Postgres client that stops reading (a big result it never takes) no longer blocks its session, and keeps its locks, forever: a write that waits a minute ends the connection. TCP keepalive finds clients gone without a word.
    • One address can no longer fill every connection slot with connections that never log in: at most 64 from it may be logging in at once; the rest are refused as "too many clients".
    • A new agent can't take the name of a dropped one while grants on worlds still name it: it would have had them. An agent a merge policy names as its reviewer can't be dropped until the policy changes.
    • An agent's max_worlds holds against forks sent at once (each counted before any landed, and all went through).
    • SHOW WORLDS (and the HTTP branches op) to an agent held to a tenant lists main and its tenant's worlds only.
    • A merge retried by its operation id while the first try is still merging returns the first's result, not "no such world".
    • While the checkpoint that upgrades a folder to format 8 fails (a full disk, a store down), merges into main are refused instead of landing with no ledger row; each merge tries the upgrade again.
    • With a transaction left open, commits no longer slow down as row locks pile up: the list is pruned again only once it doubles.
    • Statistics of dropped, merged and discarded worlds are freed, as a process running millions of short task worlds kept them all.
  • Attached databases: chronos attach and chronos serve say at the start when the source may lose a commit it acknowledged, and so refuses merges into main, instead of leaving it to the first merge.
  • Attached databases: a merge the source committed before a crash cut it short is made in Chronos as it was decided. Recovery no longer judges it again by a merge policy, merge checks, scope or stale reads, which could refuse it and leave the world open with its rows already in the source; nor as the agent whose merge happened to start the recovery. Its effects keep their delay, and its rows the taint the merge gave them. A merge recorded by 0.6.0, which doesn't name the world's version it merged at, is judged as before.
  • chronos refresh <folder> --check works beside MySQL, MariaDB and MongoDB too, comparing main with the source as their refreshes do.
  • ASSIGN WORLD writes the whole tree's tenant in one change: an ALTER WORLD, a discard or a fork of one of its worlds can no longer land between its checks and its writes and leave the root assigned with a descendant not. A world forked from the tree meanwhile refuses the assignment (40001, try again), and a fork that started before the assignment landed refuses itself instead of staying unassigned under an assigned world. The same statement (tenant and reason) run again after a crash cut one short finishes the worlds left.
  • CONSOLIDATE: its scratch world (_consolidate_<p>) can't be dropped while the consolidation runs. A drop between marking the children and the scratch world's merge into p, then a crash, discarded children whose rows never reached p. CONSOLIDATE given a world's id works as given its name (it found none of the world's children).
  • Strict proposals for payments (rules E7 and E8). A proposal whose world queues an effect on a payment-like channel (payment*, payout*, transfer*, refund*, or a policy's review_channels or new strict_channels) is strict: only a person approves it, within the policy's new strict_window (15 minutes unless set; then it's expired), and any commit to the world it merges into since it was proposed supersedes it, at approval or at the merge (40001, strict_target), so a payment never goes out against a state nobody approved. It merges whole into its parent. SHOW PROPOSALS shows strict and strict_until; SHOW MERGE POLICIES the two settings; EXPLAIN MERGE a strict line before anything is proposed.
Chronos DB 0.6.2 · Chronos