Design Message Translation

Problem Framing and Product Scope

A strong answer for problem framing and product scope, start from user-visible behavior before naming services. Frame real-time message translation as an overlay on existing chat—not a standalone translator app. Slack, Microsoft Teams, and Google Chat all need per-user locale rendering without breaking message ordering, edit semantics, or compliance boundaries. Success means a Spanish engineer reads Japanese posts in Spanish within hundreds of milliseconds while the author still sees their original text.

Consider edge cases: mixed-language messages with code snippets, rapid edit-delete cycles, channels where half the members disabled translation, and provider outages during company-wide announcements. Quantify impact: a 500 ms translation delay in a fast-moving incident channel may be acceptable; in a sales DM closing a deal it may not.

Operationally, instrument each stage independently so on-call can tell whether slowness originates in detection, MT inference, cache, or websocket fan-out. Document explicit degradation levels—from full translation through cached-only to original-text-only—and tie each level to customer-visible copy.

Key points

• Translation is a derived view; canonical message text stays in the sender language.
• Each recipient may require a different target locale in the same group channel.
• Latency budget must fit inside existing chat delivery SLOs, not replace them.

Why interviewers care

Message Translation interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

The failure that defines the design

The outage to narrate is translating the untranslatable. The model rewrites a code snippet into broken pseudo-code, localizes an @mention into a different word, or 'translates' a product name and a URL into nonsense — destroying the parts of a message that must survive verbatim. Its twin is the cost blowout: eagerly translating every message into every supported locale when most are read by a handful of people. The fix is the spine of the design: treat the original message as the source of truth and the translation as a derived, cached, additive projection; translate on demand per target locale and cache by (message, locale, model version); and preserve do-not-translate spans (code, mentions, URLs, proper nouns) by translating only the prose. State original-as-truth with on-demand cached per-locale translation up front, because translation is a derived-artifact problem, and the disasters are mangling protected spans and translating things nobody asked for.