Logging Framework

Prompt

"Design a logging library for a server. Code calls something like log.info('...'); the library decides whether to keep it and where it goes: console, a file, maybe a remote collector. Show me the classes and the path a message takes."

Clarifying questions

• What levels? DEBUG, INFO, WARN, ERROR. A message is emitted if its level is at or above the logger's threshold.
• Are loggers named and hierarchical? Yes: app.db.pool inherits configuration from app.db and app unless overridden, the way most production libraries work.
• Can one message go to several destinations? Yes. Each logger can have multiple handlers, and a message propagates to ancestor loggers' handlers too.
• Synchronous or asynchronous writes? Both must be possible. The interviewer will ask what happens when the file system stalls.
• Formatting? Pluggable; a default line format with timestamp, level, logger name, message.
• Structured fields? Key-value pairs attached to a record are a follow-up. Design the record so they fit.
• Thread safety? Required; many request threads log concurrently.

Requirements

Functional

• Named loggers with level thresholds inherited from ancestors.
• Handlers attached to loggers; a record flows to the logger's handlers and up the ancestor chain.
• Filters on handlers to drop records selectively.
• Formatters to turn a record into a string.
• An asynchronous handler that decouples callers from slow sinks.

Non-functional

• A disabled level (debug when the threshold is INFO) must cost almost nothing: no record allocation, no string formatting.
• Emitting a record must not block the caller on I/O when the async handler is used, and must not lose records silently under normal operation.
• Ordering within one thread must be preserved.
• Clean shutdown flushes pending records.

Numbers

• A busy server: 2,000 requests per second x 5 log calls each = 10^4 calls per second, of which 90% are DEBUG and disabled. A disabled call at 100 ns is 1 ms per second total; at 5 microseconds (allocating and formatting) it is 50 ms per second, 5% of a core. That is the argument for the early level check.
• Enabled records: 10^3 per second x 200 bytes = 200 KB per second to disk. Trivial for a file, not trivial for a remote collector with 50 ms round trips, which is the argument for batching in the async handler.