Model: none | Cost: $0 | Date: 2026-11-18 | Status: working
A 21-day hold on the same search query deleted 5 duplicate tickets in one prune. The page queue next to it had published nothing since 4 November. That is 14 days of zero, after a lifetime total of 36 pages.
The hold is doing the job it was written for. It is also the only signal that looks healthy. A dedupe gate on tickets is not a publisher.
Two queues, one false green
One Czech content site I run keeps improvement work and new pages in different queues.
The improvement queue is Search Console and SERP follow-ups: low click-through, a slipping URL, a query worth another look. Each ticket has a key. The key is the query, or the page when the complaint is a rank drop. A shared gate refuses to open a second ticket for that key until the previous change has been live for 21 days.
The page queue is articles. Images and video sit with it. The hold does not apply to them. The module docstring says so in one sentence: those jobs have their own queue, and this brake does not touch them.
On 18 November 2026 the page queue was empty. The note that recorded the prune also recorded the publishing history:
| Window | Pages published |
| Week of 2026-09-28 | 16 |
| Week of 2026-11-02 | 20 |
| 2026-11-04 through 2026-11-18 | 0 |
| Lifetime total in the note | 36 |
16 + 20 = 36. The two weeks in the note are the entire published set. Nothing in the note accounts for a page outside them. The last ship date is 4 November. From 5 November through 18 November the count is 0.
Five SERP tickets filed on 9 November were still sitting there. They had not run, so the hold had no closed change to measure against, and prune left them. The gate looked busy. The publisher had been idle for two weeks.
What the hold actually checks
The constant is HOLD_DAYS = 21. Block logic is three questions, in order:
HOLD_DAYS = 21
def block_reason(key, today, pending_open, explicit_hold_until, closed_on):
# An open ticket already covers this key. A second one is noise.
if pending_open:
return "open"
# A manual hold in holds.json wins until the date in the file.
if explicit_hold_until and explicit_hold_until > today:
return "hold"
# After a done/failed ticket, wait 21 days so Search Console can move.
if closed_on and today < closed_on + timedelta(days=HOLD_DAYS):
return "hold"
return None # allow Prune is stricter than "keep one". Pending tickets are grouped by key. If that key is inside the 21-day window, every pending copy is deleted. If it is not, the oldest pending file stays and the rest are deleted.
def prune(pending_by_key, in_hold):
actions = []
for key, pending in pending_by_key.items():
pending.sort() # oldest first
drop = pending if in_hold(key) else pending[1:]
for item in drop:
item.unlink()
actions.append(("drop", key))
return actions That is why 5 copies of one query disappeared and 5 different unrun SERP tickets from 9 November did not. The first key had already shipped a change, so the cushion ate the whole pending set. The other five had never closed, so they were not duplicates of a held key. "Has not run yet" is the reason they survived. It is not evidence that anyone will run them.
New pages never enter this function. A green prune log cannot tell you the article queue is empty.
The verdict is a different clock
After the hold expires, a separate function scores whether the change did anything. It does not schedule the next page.
MIN_IMP = 3
def effect(kind, baseline, current):
# Under 3 impressions the sample is noise, not a failure.
if current["imp"] < MIN_IMP:
return "thin"
if current["clicks"] > baseline["clicks"]:
return "yes"
if kind == "ctr" and current["ctr"] >= baseline["ctr"] + 0.005:
return "yes" # half a percentage point of CTR
if kind == "slip" and baseline["pos"] - current["pos"] >= 2:
return "yes" # two positions better
return "no" THIN_DAYS = 14 sits next to MIN_IMP and is never read. The thin verdict is only "fewer than 3 impressions". There is no 14-day branch. I shipped a named constant for a window the scorer does not implement.
The thresholds that do run are tight on purpose. A CTR move under 0.5 points is no. A rank gain under 2 positions is no unless clicks rose. That is a reasonable bar for "did this edit show up". It answers a question about one query. It does not answer "did the site publish this week".
What I found
The producer for pages is missing. The note says the planner that would refill the page queue does not exist. The worker that drains that queue is therefore idle for a structural reason, not because the 21-day hold blocked it.
The improvement side can still look operational:
- Prune dropped 5 waiting copies of one query and logged a cushion.
- Five SERP tickets from 9 November remained, because they had not run.
block_reasonreturnsNonefor any key that is not open and not inside 21 days. A brand-new query is allowed. A brand-new article is simply never offered to this function.
So the dashboard you get from the gate is "duplicates removed, fresh queries allowed". The publishing fact on the same day was "36 pages, then 14 days of nothing". Both are true. Only one of them is a content cadence.
The September week and the November week also show the same shape at a larger scale. 16 pages, then a gap the note does not fill with pages, then 20 pages in the week of 2 November, then stop. Bursts without a producer. The hold was added on the ticket side. Nobody added the thing that creates the next page when the queue hits zero.
What I'd change
Put a producer next to the hold, and make emptiness the alert. A daily check that prints pages_published_14d and pending_pages would have gone red on 5 November. The prune log would still be allowed to say "dropped 5". Those are different lines.
Delete THIN_DAYS or wire it. A constant that names a window the function ignores is how the next reader invents a 14-day rule the code does not enforce.
Keep the 21-day cushion on query keys. Search Console will not show a CTR move the morning after a ship, and filing five copies of one query is how a queue lies about progress. Just stop treating that log line as proof the site is still publishing.