Problem

ClickHouse Server (26.2+) has a built-in jemalloc web UI at /jemalloc that provides one-click heap profiling with in-browser flame graph generation. Keeper has none of this — diagnosing memory issues requires SSH access, manual jmfp 4LW commands, installing jeprof, and running flamegraph.pl offline.

Related PRs

• #97077 — Added jemalloc stats web UI and system.jemalloc_stats for ClickHouse Server (26.2)
• #85438 — Store jemalloc samples in trace_log, introduced per-thread profiling model, deprecated SYSTEM JEMALLOC ENABLE/DISABLE PROFILE
• #97218 — Added system.jemalloc_profile_text for querying heap profiles from SQL
• #98232 — Made jemalloc profiling cover all threads when enabled via config

Proposal

Add a /jemalloc web UI to Keeper's HTTP control endpoint (keeper_server.http_control.port) that provides:

• Live memory stats with auto-refresh charts (allocated, active, resident, mapped)
• Heap profile collection in collapsed/symbolized/raw format via a "Get Profile" button
• In-browser flame graph generation (reusing the same JS-based flamegraph.pl port from the server)

Since Keeper has no SQL engine, the implementation adds dedicated HTTP endpoints (/jemalloc/profile, /jemalloc/stats, /jemalloc/status) that call the jemalloc C API directly and return plain text/JSON. The HTML page fetches from these endpoints instead of sending SQL.

Out of scope

• Per-query profiling (Keeper has no queries or system.trace_log)
• Historical profiling via trace_log / automatic flush on OOM (separate effort)
• Enabling/disabling profiling at runtime (deprecated globally; uses config jemalloc_enable_global_profiler)