The path of an event
The path a single event takes from happening to being acted on
In a batch world an event waits in a table until an overnight job reads it. In a streaming world it moves through these stages in seconds, so the action can land inside the window where it matters. The branch is what happens when the stream itself falters, which decides whether the system degrades safely or blinds the operation.
01
The event happens
A transaction, a sensor reading or a click occurs and is emitted immediately rather than written to a table to be read later.
- Source
- Event stream
02
It is read as it arrives
A streaming processor consumes the event within moments, keeping just enough recent context to judge it against what came before.
- Stream processor
- Recent state
03
It is judged against a rule
The event is tested against the detection that matters — a threshold, a pattern, an anomaly — so a decision is reached now, not in tomorrow report.
- Detection
- Threshold
04
The action fires in the window
An alert, a block or a signal to another system is triggered while acting still changes the outcome, which is the entire reason for the speed.
- Alert
- Downstream system
When the stream falls behind
A streaming system has to be honest about lag and gaps: when it falls behind or an event is missing, it must signal that it is degraded rather than silently making decisions on stale data. A real-time platform that hides its own delay is more dangerous than a batch, because people act on it believing it is current.
Real-time is worth its cost only for the events whose decision expires before the next batch; for everything else, batch is the honest choice.
When speed is the point
When a number is right but arrives too late to be worth having
It is worth being clear about what real-time does and does not buy, because the market sells speed as an unqualified good and it is not. A nightly batch is not a worse version of streaming; for most decisions it is the correct choice — cheaper, simpler, easier to trust, and perfectly adequate when the decision it feeds is made on a daily or weekly rhythm. Planning, reporting, month-end, most dashboards: none of these get better by arriving a few seconds after the event instead of the next morning. Paying for streaming to serve them is spending on freshness nobody consumes.
Streaming earns its cost in the narrow band of decisions that expire. A fraudulent payment has to be stopped while it is in flight; approving it and reporting it tomorrow is not analytics, it is a loss with a chart. A queue building past its limit needs a second counter opened before it overflows, not a report at midnight that it did. A machine drifting toward failure this shift is worth catching this shift, while an intervention is still cheaper than a breakdown. In each case the answer and the window are the same size: acting after the window closes is not a slower success, it is a failure that happens to be well documented. Our job is to find those decisions honestly, build streaming only for them, and leave everything else where batch already serves it well.
Where the window is real
Four moments on an operation where waiting until tomorrow is the wrong answer
These are the recurring cases where the decision and its window are minutes wide, so a nightly answer is an answer to a question that already closed. They are also the test we apply before recommending streaming at all.
A payment that must be stopped in flight
A transaction that looks fraudulent is being authorised right now, in the second it takes to clear.
The stream scores the transaction against the pattern as it arrives and can block or hold it before it settles, not report it after.
- Transaction stream
- Scoring
- Hold or block
A queue about to overflow
Arrivals at a counter, a gate or a call line are building past what the current capacity can clear.
A live count against a threshold signals for another counter to open while the queue can still be recovered, not at the end of the day.
- Live count
- Threshold
- Staffing signal
A machine drifting toward a fault
A reading is trending out of its safe range during this shift, before any scheduled inspection.
The stream watches the trend continuously and alerts while an intervention is still cheaper than the breakdown it prevents.
- Sensor stream
- Trend detection
- Maintenance alert
A shelf or stock that just ran out
An item has gone to zero at a location and every minute it stays empty is a sale not made.
The event triggers a replenishment signal the moment stock hits the floor, rather than a report that surfaces the gap the next day.
- Stock event
- Reorder trigger
- Fulfilment
Our role
What we stream and what your operations team decides on
- 01
The honest scoping first
We start by separating the decisions that genuinely expire in minutes from the ones a daily report already serves, so streaming is built where the window is real and not sold as a default.
- 02
A stream that admits its lag
We build the pipeline to signal when it is behind or missing events rather than deciding silently on stale data, because a real-time system that hides its own delay is worse than the batch it replaced.
- 03
The action stays your call
We put the detection and the signal in front of your operators and, where you choose, wire it to act automatically — but which events warrant a block or a callout is your rule to set, not ours to assume.
Which decisions are worth the cost of acting in seconds, and what the operation does when an event fires, are your calls to make; we build the streaming that reaches those decisions inside their window and is honest whenever it cannot.
Common questions
What buyers ask
How do we tell which decisions genuinely need real-time and which do not?+
By asking one question about each decision: does the window in which acting changes the outcome close before the next batch would run. If a fraudulent payment has to be blocked in the second it clears, or a queue needs another counter before it overflows, the window is minutes and batch is too late — real-time earns its cost. If the decision is made in a morning meeting, at month-end, or on any daily or weekly rhythm, then an answer a few seconds after the event is worth nothing more than one that arrives overnight, and streaming is spending on speed nobody uses. Most requests for real-time, examined this way, turn out to be a handful of genuinely time-critical decisions surrounded by many that batch already serves well. We map that honestly before building, because the wrong reason to stream is that the data could be fresher, and the right one is that a decision expires.
Is streaming analytics much more expensive to run than a nightly batch?+
It is more expensive, and the honest way to think about it is per decision rather than in the abstract. A streaming platform runs continuously and processes each event as it arrives, so it costs more to operate and more to build correctly than a job that runs once a night — there is more to keep alive, more to monitor, and more that can go wrong at three in the morning. That cost is easily justified for a decision that expires in minutes and is worth real money each time, and impossible to justify for a report someone reads over coffee. This is why we scope the time-critical decisions first and build streaming only for them, keeping everything else on the cheaper batch it is already well served by. The mistake to avoid is streaming the whole estate because part of it needed to be fast.
Can real-time analytics run on the operational systems we already have?+
Often yes, and the first step is checking whether your systems can emit events rather than only being queried in bulk. Real-time analytics needs the source to announce that something happened — a payment, a reading, a scan — as it happens, instead of writing it to a table that a batch job reads hours later. Many modern operational systems can already publish these events, or expose a log we can read as a stream, in which case we build on what you have. Where a system only supports periodic extraction, we are honest that it cannot feed a true real-time decision without a change to how it emits data, and we would rather tell you that than layer a streaming platform on a source that fundamentally updates in batches. What we do not do is claim real-time on top of a system that can only be polled, because the result would look live and act stale.
In this cluster
Next step
Scope it with us
A working session that separates the decisions that expire in minutes from the ones a daily report serves fine, and sizes streaming only where the window is real.

