advanced 2 min answer

To hit a launch date, a 60-engineer marketplace organisation of eBay's shape stops all technical debt work for two quarters. Nothing is deleted, no dependency is upgraded, no flake is fixed. Assume the launch succeeds. What degrades, in what order, and what becomes genuinely irreversible?

ebaytechnical debtdependenciescimigration cost
Show the full answer Hide the answer

Weeks 1 to 4: nothing

This is the politically important part. The absence of visible harm in the first month is read as evidence that the debt work was never necessary, and that reading is what makes the freeze get extended. Anyone arguing against the freeze needs to have said in advance that this is what the first month looks like.

Weeks 4 to 10: the pipeline degrades first

Nobody is deleting tests or fixing flaky ones, and every shipped feature adds tests to a suite that is no longer pruned. Pipeline duration and re-run rate are the first measurable casualties, and both are leading indicators because they feed everything downstream.

Weeks 8 to 16: change size rises, then change failure rate

A slower pipeline raises the fixed cost of a submission, so changes get batched to amortise it. Bigger changes are harder to review and harder to bisect, and change failure rate follows within a few weeks. By now the freeze is producing the outcome it was meant to avoid: slower delivery.

Weeks 12 to 20: dependency drift stops being tidiness

A framework two minor versions behind is an afternoon. Four minors and a major is a project, because breaking changes compose: each upgrade's migration guide assumes you came from the previous version. This is where the cost curve turns superlinear, and it is the first thing that starts to feel irreversible.

It is also where it stops being an engineering preference. A dependency with a published vulnerability that you cannot upgrade without a two-week project is a security finding with a clock on it.

Weeks 16 to 26: knowledge concentrates

The workarounds introduced under deadline are now load-bearing and undocumented. Onboarding slows because the explanation for why something works lives in one person's head, and you find out how concentrated it is when that person leaves, which is disproportionately likely after a two-quarter crunch.

What is genuinely irreversible

A data model change deferred past the point where the data outgrew it. A schema migration that is one hour at 10 million rows is a multi-week online migration with backfill, dual-write and verification at 2 billion. The cost never comes back down, and no amount of later investment restores the cheap option.

What to protect even during a freeze

Three things, and only three, because their cost curves are the superlinear ones:

  1. Security-relevant dependency updates.
  2. The flake budget, because it compounds into everything else.
  3. Any change whose cost grows with data volume, above all schema changes.

Everything else genuinely can wait two quarters, and saying so is what makes the three credible.

When this is the wrong analysis

When the system's remaining life is shorter than the debt's payback period. A campaign site with a six-month lifespan, or an acquired platform already scheduled for decommissioning, should accrue debt deliberately — paying it down is spending on a future that will not arrive. The mistake is applying that reasoning to a system everyone privately expects to still be running in five years.