flintspark Book a demo

Workflows that wait and check first

A flintspark workflow is a sequence of skills, checks, questions and approvals that delivers one business outcome. It runs durably: it can wait days for a person, resume after a restart, and check whether an irreversible step already happened before it redoes it. When it cannot tell, it asks.

  • Read the weekly pricesDone · a skill
  • Which region takes the extra volume?Waiting for a person · day 2
  • Submit the supplier requestNeeds approval, by default
A sketch with the energy demo's names, not a screen from the product.

What is a workflow?

The steps that turn one request into one outcome, such as a procurement recommendation. Some steps are skills Flint already has. Others check a value against a rule, ask a person a question, or wait for an approval.

In the platform the workflow graph shows it as a picture: the steps, the branches, the approval points and the report at the end.

A sketch with the energy demo's names, not a screen from the product.

How does it wait days for an answer?

When Flint needs a person, the workflow opens a question or an approval in the platform's inbox and stops. While it waits it uses no compute, and the wait can last days.

  • The answer arrives, and the workflow carries on from the same step.
  • Each wait has a time limit set in configuration. If an approval wait times out, the run is marked expired and nothing is sent.
A sketch, not a screen from the product.

What happens after a restart?

The workflow resumes from its last checkpoint. If the server restarts while a run is waiting for a person, the run is still waiting when the server is back, and the answer is not lost.

Steps that only read, or that can be undone, are safe to run again. Steps that cannot be undone follow the check in the next question.

A sketch, not a screen from the product.

How is a double send avoided?

Before an irreversible step is sent, the workflow writes down that it intends to. An irreversible step is sent once and never retried on its own.

If the server crashes mid-step, recovery runs a post-condition check first: does the request now exist on the site? Only a clear "no" allows a retry, and the retry needs a new approval. If the check cannot tell, the step is marked uncertain and a person decides: it happened, it did not happen, or cancel.

How are watches scheduled?

A watch is a workflow that checks its sources on a timetable and alerts a person when something changes. A schedule starts each check, so nobody has to remember to run it.

A watch can compare a source's last-updated time with the one it saw last time and skip a run when nothing has moved. It can also hand a finding to a person for approval, like any other workflow.

A sketch, not a screen from the product.

Questions about this page

What happens to a workflow if the server restarts?

It resumes from its last checkpoint. A run that was waiting for a person is still waiting when the server is back, and the answer is not lost. Steps that cannot be undone are sent once and never retried on their own.

How does a workflow avoid sending something twice?

Before an irreversible step is sent, the workflow records its intent. After a crash, a post-condition check asks whether the step already happened. If the check cannot tell, the step is marked uncertain and a person decides. A retry after a proven miss needs a new approval.

What is a watch?

A watch is a workflow that checks its sources on a timetable and alerts a person when something changes. Like any workflow, it can ask questions and wait for approvals, and it can skip a run when the source has not changed since the last check.

See Flint run

Bring one workflow you'd rather not do by hand. We'll show Flint learning one like it, and stopping before anything it can't undo.

We reply from hello@flintspark.ai.