Katkat
Back

Should We Launch Group Calling?

MetaData Scientist · Technical Screen
HardData ScientistTechnical ScreenAnalytics & Experimentation

Last updated

A Meta Data Scientist technical-screen case: using only one-to-one call logs, decide whether to build a group calling feature and how to measure a launch's success. It tests product analytics, proxy-metric design, network analysis, causal inference, and experiment design under social-network interference.

Question

You work on a consumer calling product (think Messenger/WhatsApp-style voice) that currently supports only one-to-one voice calls. The team is considering whether to launch a group calling feature. You have access to the existing call logs and a daily-active-users table:

Table: users

  • user_id BIGINT
  • country_code STRING

Table: calls

  • call_id BIGINT
  • caller_id BIGINT (references users.user_id)
  • recipient_id BIGINT (references users.user_id)
  • started_at TIMESTAMP
  • ended_at TIMESTAMP
  • status STRING (e.g. completed, missed, failed)

Table: daily_active_users

  • activity_date DATE
  • user_id BIGINT
  • country_code STRING

Assume one row per call attempt, and that all date-based analyses use the Europe/London timezone.

Using only these existing one-to-one call logs (no group-call feature exists yet), answer the following product analytics questions:

  1. How would you determine whether the product should build a group calling feature? (i.e. is there unmet demand?)
  2. If group calling is launched, how would you decide whether the launch was successful?

In your answer, cover:

  • how to infer latent/unmet demand for group calls from one-to-one behavior
  • what proxy metrics you would build from call sequences and user networks
  • alternative explanations and confounding factors (selection bias, spam, network quality)
  • what success means across adoption, engagement, quality, and retention
  • the difference between launch (novelty) metrics and long-term value metrics
  • how you would design an experiment, including randomization challenges from social-network interference
  • guardrail metrics and failure cases
  • how you would interpret mixed or ambiguous results