Skip to content
Henrique Senra

PayPal · Shipped Q3 2025

Counting about 130 workaround configurations to get one missing platform feature built

PayPal's personalization platform had no IN operator, so teams worked around it with an extra call on every load. I found the gap, counted the workarounds and made the case on a platform that was preparing for a possible sunset.

~130

workaround configurations counted across Home and other domains

  1. 01About 130 workaround messages counted across Home and other domains, with a developer's latency study.
  2. 02Operator shipped in Q3 2025. Home adopted it at once and retired a few dozen messages plus the objects that supported them.
  3. 03No product metric and no latency figure after the change.
  4. 04Cost and latency removed, directional.

TL;DR

  • Problem. PStudio’s contextual rules, which decide who sees a message, had no IN operator for text values. Teams worked around it with an extra call to the experimentation tool on every load.
  • My role. I found the gap, scoped the cases and showed that it was a real need that affected the business.
  • Decision and cost. I argued from a count of about 130 messages and a developer’s latency study to a platform team that wanted to skip it. The feature was a small side task.
  • Outcome. The operator shipped in Q3 2025. Business end: cost and latency removed, directional. Nothing was measured afterward.

Context

PStudio is PayPal’s internal no-code platform for personalizing the app. Its contextual rules combine conditions, such as country and app version, to decide who sees a message. For a long time those rules had no IN operator on text values, which checks whether a value is in a list.

The gap hurt during the rollout of the new Home experience. It launched first in the United States, then in a few European countries, then in the top 10 countries where PayPal operates, with an expansion to the top 50 planned. Each country received the new version from a specific app version, such as 8.34, then 8.38, then 8.42. The ramp ran through Elmo, PayPal’s experimentation tool. When the ramp reached 100% in the ten countries, the plan was to unplug the experimentation tool and use PStudio’s native rules to control the versions.

Without IN, the team would have needed one rule set per country, and there were already 10 to 12 countries. Some, such as England and Germany, had an intermediate version, first a navigation change and then a dashboard change. Those countries needed a range rule, such as one version below 8.50, another between 8.51 and 8.58, and a third above 8.58. That rule is almost impossible without IN, because it has to combine a list of countries with a range of app versions. The logic ran on the main dashboard and the navigation, so it covered literally all users.

Other teams used the same workaround, some through RPS, the real-time segmentation service, and most through Elmo. RPS was not built for real-time device checks such as app version. The result was dozens, maybe hundreds, of PStudio objects making unnecessary calls to Elmo. That added latency, infrastructure cost and load on the experimentation tool. A minor effect was that the tool kept returning positive, so users were marked as treatment in experiments already at 100% with no control group.

What I owned

I found the gap, designed the fix, scoped several cases and showed it was a real need that affected the business. In the team’s account, “we” noticed it, and in my first recall it was me. I answered “I did” when asked directly. I documented the impact at the time, but I no longer have that document.

What belonged to others. The platform team built the feature. A developer on our team ran the study that showed how much the dashboard latency grew.

Options

flowchart LR\n  A["Rule needs: country in a list and app version in a range"] --> B["Before: extra Elmo call or one rule set per country"]\n  A --> C["After Q3 2025: native IN operator in the rule"]

The platform’s team did not want to spend time on it. The tool was preparing for a possible sunset a year or a year and a half later, and the request was not a priority. The record shows no list of alternatives I weighed. What it shows is the choice between keeping the workaround and asking for the feature, and what moved the platform team.

Option Cost
Keep the workaround An extra Elmo call on each load in every affected message, plus one rule set per country.
Ask for the operator A hard negotiation with a team winding down, and a feature that needed tests and test cases that an engineer said might be more work than the change.

We counted what we could see. Dozens of messages in the Home team alone, some dozens more from other domains we knew of, for a total of more than 100, about 130 messages. There was no way to search all usage across the company, so the count is a lower bound for what we saw. The count, the number of calls in PStudio and Elmo, and the developer’s study convinced them. Latency is a guardrail metric because it affects bounce and checkout rates. In the end it was not a prioritized criterion, but it forced the team to accept the request. They did not have a clogged backlog, so it was a matter of opening the backlog and letting it happen.

Results

When IN shipped, in Q3 2025, Home adopted it immediately. We retired not only the workaround but some objects that existed to support it. In our domain we corrected a few dozen messages, which made work easier. I cannot say what other teams did.

There is no specific product metric. We did not monitor latency at the time, because the concern was retiring legacy configurations to prepare for a new app version. I described it in my first recall as the kind of small change that saves a few milliseconds per call, in a product that makes hundreds or thousands of calls across millions of daily views, which I believe is a real business effect. I have no business figure to show it.

Friction and what I would change

Adoption had no problems. It was more a discovery of a small change with a significant impact that was implemented very quickly. It also covered a technical debt, since IN is too common an operator to be missing.

I would do one thing differently. I would monitor the impact at the end of the chain. This was a side quest of a bigger task and a quick win that could have been better monitored.

Limits

  • The count of about 130 is approximate and covers what we could see. It is not the number of messages retired.
  • Home retired a few dozen messages. What other teams did is unknown to me.
  • No before or after latency or cost was measured. Any business effect is directional.
  • Names of colleagues stay private.

Related cases