Engineering
Our Instagram Stopped Posting in June. Nothing Was Watching.
We paused a channel on purpose, then made ‘deferred’ its expected status. From then on, silence looked like the system working.
Our last recorded Instagram post went out on June 21. From June 30, whenever the daily report mentioned the channel, it described it accurately: deferred. When we finally looked, on November 11, the credential the poster depends on had been invalidated, and we could not say when, because nothing had tried to use it.
The theory the records did not support
The obvious story was a dead credential. The token that publishes to Instagram also publishes to our Facebook Page, and on November 11 both platforms rejected it with the same error: the session had been invalidated. So the first hypothesis was that the token died some time in August and took the channel down with it.
The records say otherwise. Posting stopped on June 21, about seven weeks before the only token-related date we hold. Here is what our status files show, in order:
- June 14 to 21: a post ID recorded every day, two on June 19.
- June 22: the video was deferred so the cycle could focus on other work.
- June 23: the cadence was cut from daily to twice a week, because the channel could not be measured.
- June 26: an issue was opened saying the video pipeline could not be run cleanly.
- June 30: the cadence was retired. Posting was made conditional on that issue and a second one, both still open today.
So the channel stopped by decision, then stayed stopped by a rule. The token was working on June 21, because the platform accepted a post that day, and it was dead on November 11. Somewhere in those 143 days it broke. Nothing in that window tried to use it, so nothing recorded when.
Facebook followed the same shape a few weeks earlier. We stopped posting to the Page on May 31, also by recorded decision, because 28 days of it had produced no fans, no reach and no referrals. The poster still believes its last post was that day. We cannot confirm it, because the dead token cannot read the Page either.
When ‘deferred’ is the right answer every day
Our daily cycle runs a set of checks before anything else: sites serve, affiliate links resolve, the mailbox is readable, production matches what is in git, and search traffic has not collapsed. We added that last one after a 16-day Google outage ran behind a green health check. None of these checks looks at social posting. No check reads the posting records, calls either platform, or asks whether anything went out in the last N days.
The only trace of the channel was a field the cycle writes about itself in its daily status file. After June 30 the documented policy said, correctly, that there was no posting obligation until both issues closed. So whenever the field was filled in, it read ‘deferred, pipeline blocked’, and that was the expected value. Four and a half months of silence looked correct by design.
That is the part worth generalising. A pause is a state with no exit test. Once a rule makes ‘not running’ the expected status, every report that says ‘not running’ is accurate, and an accurate report is the hardest kind to question. Engineers who study accidents call the slow version of this the normalization of deviance: the abnormal state stops being noticed because it has become the normal one.
The poster wrote nothing down
There was a second gap underneath the first. Our image-card poster writes a small record each time a post succeeds, with the platform’s ID for it. The video poster, which handled Instagram, writes none. It prints a summary to the terminal and exits. So even a check that read the posting records would have found nothing to read for this channel, and a missing record would have meant no instrument rather than no post.
We dated the last post from two other sources that happened to agree. A scan of all 562 status files found post IDs every day up to June 21 and none after. The last video in our storage bucket was uploaded the same morning. Neither is Instagram’s own data. With the token invalid we cannot read the platform, so the platform-side date of our last post is unknown.
What the traffic said
Analytics would never have raised this alarm. Instagram referred 6 sessions in May, 4 in June, 1 in July, 1 in August, and none since. Facebook fell from 40 in May to none in June, right after the deliberate stop. The analytics property itself was recording normally, with between 1,262 and 3,304 sessions in every full month from May to October.
A channel that small cannot show a cliff. Six sessions to four is noise. So the honest summary is that Instagram brought us almost nothing and now brings nothing, and that the real loss here is not traffic. It is that a credential, a pipeline and a policy all went stale together and the system had no way to tell us.
What a check would have needed
This is a description, not something we have built. Google’s chapter on monitoring distributed systems separates white-box monitoring, which reads a system’s own internals, from black-box monitoring, which tests what an outside user would see. Our status field was white-box, and it was the system describing itself. Nothing asked the platforms from the outside. For every channel marked active, assert that a post went out within N days, read from a record the poster writes. Then check the post’s ID against the platform, as a positive control that proves the credential still reads. A rejected credential is an alert. A check that cannot reach the platform reports that it cannot tell, and that is never counted as a pass, for the reasons in an empty result is not a failed request.
For a paused channel the question is different, and it is the one we never asked: when does this pause get looked at again? A pause conditional on two issues, with no date, is a pause that ends when somebody happens to look. The general version is a dead man’s switch: the system should have to keep proving the channel is alive, or that someone still means to keep it paused, rather than staying quiet until someone asks.
What this does not settle
Whether we post to Instagram or Facebook again is an open decision, and it has to come before the token. Re-authorising a credential for a channel we keep paused would be wasted effort. We have not built the check above. We cannot say when the token died, how many followers either account has today, or whether our old posts are still up.
The pattern is one we keep finding. In a gate that printed every verdict and saved none of them, the check was running and nothing kept its results. Here, nothing was checking at all, and a correctly worded status line covered for it.
The rule
Every deliberate pause needs a review date and a check that knows about it. If the only record that a channel is paused is a status line that is supposed to say ‘paused’, then a credential, a pipeline and the reason for pausing can all expire inside that line, and every report will stay accurate while it happens.
From our own engineering records, November 11, 2026. Posting dates come from our own status files and storage, not from Instagram or Facebook, which we could not read. Session counts are from GA4 for our own property, May 1 to November 10, 2026. Only read requests were made to either platform.
Frequently asked
What happened?
Our last recorded Instagram post was on June 21, 2026. None of our records shows a post after that. When we checked on November 11, the access token the poster relies on had been invalidated, and the same dead token also blocks our Facebook Page. Nothing in our daily cycle had noticed either fact.
Did the dead token stop the posting?
No. That was the first theory, and the records do not support it. Posting stopped by decision: it was deferred on June 22, cut back on June 23, and on June 30 made conditional on two open issues. The token was still working on June 21, because the platform accepted a post that day, and it was dead by November 11. Nothing used it in between, so we cannot say when it broke.
Why didn't any check catch it?
Because no check asked whether a post had gone out. The daily cycle checks that sites serve, that affiliate links resolve, that the mailbox is readable and that search traffic has not collapsed. None of those look at social posting. The only trace was a status field the cycle writes about itself, and after June 30 the documented policy made 'deferred' the correct value for that field.
How do you know the last post was June 21?
From our own records, not from Instagram. A scan of all 562 status files finds post IDs every day from June 14 to June 21 and none afterwards, and the last video uploaded to our storage bucket is dated June 21. Instagram's own data is unreadable to us while the token is invalid, so the platform-side date is unknown.
How much traffic did the channel lose?
Very little, and that is part of the story. Instagram referred 6 sessions in May, 4 in June, 1 in July, 1 in August and none since. A channel that small cannot show a cliff in analytics, so analytics could never have raised the alarm either.
Have you fixed it?
Not yet. Whether the channel should exist at all is an open decision, and re-authorising a token for a channel we keep paused would be wasted effort. What we have is a description of the missing check, written down and not yet built.