Engine v0.35.0

Changed: the Overview (#117)

• A window of 24 hours, 7 days or 30 days, from three links at the top
  (?window=). The chart is per hour over a day and per day over a week or a
  month; a day's bar opens that day's runs.
• Runs that needed a retry are counted: a failure that healed, which does
  not show as one.
• The next trigger is in the strip; the eight soonest sit beside the chart.
• The head says what runs, what retries and what waits in the queue, each a link
  to its list. "In progress" shows only when something does.
• The donut goes; its four counts are in the strip.

No migration. No configuration changes.

Engine v0.34.0

Changed: a workflow's page (#116)

• Pause and Resume sit beside Run on a workflow with a schedule, one click,
  as on the Workflows list.
• When it runs next is a number in the strip: "in 12m" and the UTC time,
  "paused" or "manual".
• A one-step workflow shows its step, not a canvas holding one box: its
  command and what it runs in, beside its year of runs.
• The head says, in one line, whether the schedule runs, the project, the cron,
  the size of the graph and the tags.
• The workflow's runs no longer repeat its name on every row.

No migration. No configuration changes.

Engine v0.33.0

Changed: a load's page (#115)

A load's page answers whether a destination is being fed and how it is
configured, now in the first screen.

• The configuration is in tabs: Summary, Sources, Transform, Destination
  and Schema. A from.Many's settings that every source shares are said once
  ("Common to 10"), with a table of only what differs between them. A link to
  #sources opens on that tab.
• Rows read beside rows written. The newest load says "0 of 370 read" when
  its run's stages know, in red when it wrote nothing of what it read. A
  mapped step, whose stages belong to one instance, shows no count.
• The preview hides the columns that are empty in every row it returned,
  with a switch to show them and a note naming them. It still loads only when
  asked. The SDK's own columns come last.
• The recent loads sit beside the page; the head says how the load stands,
  what it is, who writes it and when.
• Freshness is a dot and a word on every page, like a run's state.

No migration. No configuration changes. A pipeline describes its
configuration from SDK v0.85.0; a load written by an older one shows none.

Engine v0.32.1

0.32.0 was tagged and never published: its release stopped before any
image was built, on a manifest that pinned the gateway at the engine's version.
0.32.1 is the same engine with that pin back at the gateway's own.

Changed: the run page (#114)

The run page is where an operator lands from an alert, to find out whether it
worked, where it stopped, and how to run it again.

• Rerun, beside the title. A workflow with no params reruns in one click;
  one with params opens the Run… drawer filled in with this run's values. A
  param the workflow no longer declares is named and not sent. It is a manual
  run: for a run with a slot, the page says the rerun reads now's window, not
  the slot's.
• A one-step workflow shows its step, not a canvas. The step's command,
  runtime, SDK and duration on one line, and each SDK stage (check, source,
  map, target) with the numbers the graph showed. It follows a live run.
• Step output is a table: one row per attempt, opening to its error and
  log; a mapped step opens to its instances.
• Alerts are one line each, the error opening under its line.
• The top bar names the run (daily_sales · 2c7233f8), and its id copies.

No migration. No configuration changes.

Engine v0.31.0

Changed: a denser console (#113)

The console spent its space on ceremony -- large numbers, caps eyebrows, tall
table rows, full timestamps in every cell. The shared components now take the
compact measures, and every screen shows more with nothing removed.

• Tables have 34px rows and a 30px sentence-case header. Dates in a table
  read Oct 16 02:23 (UTC), with the full date on hover. The Runs list shows
  about 20 runs in a 1440×900 window, up from 12.
• One panel heads every block (title, count, actions), replacing the caps
  section titles. The page title is smaller and the padding tighter.
• A state is a dot and a word, not a badge. A retried run shows ↻N.
• Key numbers sit in one strip, the Overview's four cards included.
• Small charts are smaller: the load bars are 44px tall, and the year of
  runs uses 8px cells.
• The run page: the phases are a timeline, the error is a line you open
  for the full text, and the auto params are a key-value grid.

No migration. No configuration changes.

Engine v0.30.0

Fixed: no flicker between screens on real infrastructure (#112)

On a deployed console the screens took a moment to change and at times
flickered; the sidebar's search, especially, disappeared and reappeared.

• Nothing pops in after the first paint. Controls that need a script --
  the search, Copy, Live, tab bars, group toggles, the SQL workbench's strip
  and recent rail -- were drawn hidden and revealed by the deferred script, so
  a paint before it drew the page without them. One inline line at the top of
  the head now marks the page before anything is drawn, and they are there
  from the first frame. A stream's #deadletter link opens on Dead letter.
• A page the pointer rests on is fetched before the click (speculation
  rules, prefetch only). Never sign-in or sign-out, the API, the assets, or a
  Preview link -- that one is a billed warehouse query.
• A click is acknowledged: a thin bar at the top when the next page takes
  more than 120 ms.
• The swap fades through: the old page leaves, then the new one arrives,
  never both half-visible at once. The sidebar and the top bar hold still.
• The graph comes with the page instead of a round trip later, so the
  workflow and run pages no longer draw an empty canvas first.

No migration. No configuration changes. If the screens still feel slow on a
deployment, check that the proxy keeps Cache-Control: immutable on
/assets/*?v= -- the console sends it, and a proxy that rewrites it costs a
round trip per asset on every screen.

Engine v0.29.0

Changed: Catalog becomes Streams and Loads (#111)

Catalog listed two mechanisms as "destinations", and a stream's dead letter as
a destination of its own. The console now has an Ingest group with two
screens, each answering one question:

• Streams (/ingest/streams): what comes in through the gateways. One card
  per stream, POST <path>, with its compact flow and its dead letter hanging
  off the destination ("refused batches"). A stream's page has Overview,
  Configuration (#config, everything it declares, "—" for what it does not)
  and Dead letter (#deadletter: the path, the gateway metric that counts what
  it buried, the ls/resend commands), plus Send test event, a drawer with
  a validated body and a curl to copy. The console sends nothing.
• Loads (/ingest/loads): what the workflows wrote, grouped by workflow,
  with freshness, last load and rows. A load's page is one step on one
  destination: rows per load (each bar opens its run; zero is red), recent
  loads, configuration, Preview and Query in SQL.

The sidebar is now Pipelines · Ingest (Streams, Loads) · Consume (Endpoints,
SQL) · Workspace, and ⌘K lists streams and loads instead of destinations.

Old links keep working. /data answers 301 to /ingest/loads;
/data/target?u=<uri> answers 301 to the stream whose sink or dead letter the
URI is, else the load that writes it, else /ingest/loads. /data/endpoints
is unchanged.

No migration. No configuration changes. The streams list reads every stream's
configuration in one query (gateway_streams joined to writer_configs).

Engine v0.28.0

Changed: the Pipelines screens (#110)

Workflows. The search answers while you type: the server still searches
(the list is paginated), and the list is swapped under the box. State and
schedule are Segmented filters, each choice with its count. Tags are a popover
with a box that filters them, and two tags combine with AND (?tag=a&tag=b).
Each row draws its last 14 runs as squares linked to the runs, read in one
query for the page, and says its schedule in words for the shapes that are
certain (0 4 * * * → "every day at 04:00"); any other expression is shown
as written.

Workflow. The title is the slug, with its project, schedule in words and
Active/Paused above it, and the four numbers are one Stats strip. A workflow
with parameters runs from a "Run…" drawer with a preview of the payload; the
trigger is the same POST. With no script, the form shows in place.

Runs. The state filter counts each choice, each duration has a bar against
the page's longest, and a "Live" switch (off by default, remembered per
browser) refreshes the list every 5 s, washing the new runs. Nothing is
requested while the tab is hidden.

Run. A strip of phases under the header -- scheduled or triggered, how late
it started (from the slot, like delay_seconds), how long it ran, how it
ended -- refreshing with the rest of the live head.

No migration. No configuration changes. On a large history, runs would
benefit from an index on (workflow_slug, criado_em) (#109).

Engine v0.27.0

Added: a destination's page draws how data reaches it (#108)

Each described writer's configuration card on /data/target now opens with
its flow: the source (POST <path> and its format for a gateway stream, the
driver and what it reads for an SDK pipeline), the buffer when its flush is
configured, the destination, and the dead letter as a dashed branch -- where
refused batches go.

It draws only what the configuration declares. Every label is a configuration
field or the destination itself, and a writer that did not describe itself
(a Python, dbt or shell step) gets no diagram. Drawn on the server, it needs
no script, and below 700px the nodes stack.

The site's illustrations of the console now use the console's own palette.

No migration. No configuration changes.

Engine v0.26.0

Added: ⌘K goes anywhere in the console (#107)

Press ⌘K (Ctrl+K) on any signed-in page, or the sidebar's "Search or jump
to…", to search the sections and everything the console lists: workflows,
projects, destinations and endpoints. Pasting a run's id offers that run.
↑↓ choose, ↵ goes, ⌘↵ opens a new tab, and Esc closes the palette and
gives the focus back. It works from inside the SQL editor too.

It only navigates; nothing in it changes the system. Matching is a plain
substring, ranked by the label's start, then a word's start, then anywhere.
The index is GET /api/palette -- behind the session like every page,
no-store -- fetched once per page and filtered in the browser.

No migration. No configuration changes.