Pipelies is a field guide to the ways a stitched-together data stack hands decisions numbers that are no longer true: seven suspects, fourteen scenarios you can step through, and an interrogation for your own pipeline.
TL;DR: Pipelies is a field guide to the ways a stitched-together data stack hands decisions numbers that are no longer true. Seven suspects, fourteen lies, each one a scenario you can step through, plus a six-question interrogation for your own pipeline. See the lineup.
Job status: succeeded. 412 rows never made it through.
Replica healthy. It's several seconds behind, because it's lunchtime.
Risk score: low. Computed at 2 a.m. Someone else took over the account at noon.
Each of those statements came from a system working as designed, and each is accurate about itself. None of them tells the decision reading the number whether that number is still true.
Pipelies is our field guide to those statements. We built it so you can watch one happen.
Each lie gets a page that opens with the same three facts: what the system says, what's true, and what it cost. Then you step through the scenario moment by moment, with what the decision saw next to what was actually true.
| Suspect | Its alibi | Lies on file |
|---|---|---|
| Warehouse | Correct as of the last load, which may have been last night | The Overnight Truth, The Averaged Lie, The Next-Day Alert |
| Read replica | Usually a moment behind the primary, and further behind when you're busy | The Vanishing Order, The Peak-Hour Lag |
| Stream job | Keeps the running totals, and at the busiest moments they trail the events | The Scattered Balance, The Crowded Total |
| Batch ETL job | Marks the run green even when some of the rows didn't make it | The Silent Skip, The Late Arrival, The Linked Accounts |
| Feature store | Hands your models features from the last scheduled run | The Frozen Score, The Stale Segment |
| Agent memory | Holds whatever was true when the agent last checked | The Agent's Alibi |
| Metrics layer / BI | Ask two dashboards for revenue and you may get two numbers | The Two Truths |
Each scenario plays out on a timeline. Here's the last moment of The Frozen Score: the customer's risk is really high, the feature store still says low, and the transfer check approves the request.
One click runs the same moments again with Tacnode in place of the stale read. The events don't change. The check sees what happened in the last four minutes, and the transfer is held.
Four more tabs sit behind every scenario: where the lie shows up in other workflows, why the system gets it wrong, the fix, and the questions to put to your team. The fix tab draws what comes in, what stays where it is, and what the decision gets.
The Frozen Score. At 2 a.m. the nightly job scores a customer low risk. At 12:10 someone changes the account's email and phone number. At 12:12 they add a new payee and raise the transfer limit. At 12:14 they request a large transfer to that payee. The check reads the 2 a.m. score and approves it.
The Scattered Balance. At the open, a customer with $5,000 buys $3,000 of event contracts. Two seconds later they place a $4,000 crypto order. The fill is on Kafka and the stream job hasn't folded it in yet, so the order check sees $5,000 of buying power. The customer had $2,000.
The Agent's Alibi. Two agents read the same $5,000 budget a couple of seconds apart. Agent A commits $4,000. Agent B, acting on what it read, commits $3,500. The budget is $2,500 over, and there was nobody between the read and the action to notice.
A decision reading a stale number gets a fast, successful response. There's no error, no timeout and no alert, because nothing failed.
Of the fourteen lies, nine are numbers that were true earlier, three are missing something that should be there, and two are systems that disagree with each other. Outdated, incomplete and inconsistent are the three ways a context gap shows up.
None of these systems is bad at its job. A warehouse answers questions about closed periods very well. The trouble starts with the copying: each hop between where a fact lands and where a decision reads it adds its own delay, and reports only on itself.
So keep your streams. Lose the pipelines that copy them into caches.
Six questions: how often the data your decisions read gets refreshed, how many systems one decision reads from, what happens when two of them disagree, how you learn a job dropped data, whether automated actions read that data, and what acting on a stale number costs you.
You get a lie score out of 18 and a profile: Mostly Honest, Chronic Staler, Two-Faced or Pathological. Take the interrogation.
Tacnode keeps the numbers your checks depend on, such as exposure, buying power and limits across accounts, consistent and current at peak, without copying them through pipelines into caches. Source data arrives from the systems you already run. A decision reads its context and produces a verdict, a hold or an action, and that output is written back as context for the next decision.
Every lie page shows what would change for that decision and what stays where it is. If one sounds familiar, bring that decision to a demo, and we'll trace what it reads and how old that data is when it runs.
Former Cockroach Labs. Tells stories about infrastructure that actually make sense.
Fresh Context Newsletter
The argument behind the posts — what breaks when AI acts on stale data, from the team building the Context Lake. About twice a month.
Book a demo and discover how Tacnode can power your AI-native applications.
Book a Demo