run: cap Redis maxmemory to N% of total_system_memory before init_commands

[fcostaoliveira]

Summary

Adds an opt-in mechanism to set Redis maxmemory proportionally to the bench-server's total memory before any spec-defined init_commands run, so workloads that previously OOM-killed the OS now hit Redis-level OOM errors instead.

Three sources, in precedence order:

Source	Where
--no-maxmemory-cap	CLI — force-disable for this run, beats everything below
--maxmemory-pct N	CLI — override applied regardless of spec
dbconfig.maxmemory_pct: N	spec field — per-spec opt-in

When none is set, nothing is sent. Existing specs and existing CI pipelines (RediSearch, RediSearchEnterprise, anyone else consuming this CLI) see no behaviour change unless they explicitly opt in.

Why

The search-expire-doc-10-milliseconds spec on RediSearch makes the bench-server (m7g.8xlarge / m7i.8xlarge) unreachable via SSH around 140-170s into the test, with paramiko: Error reading SSH protocol banner after Connection closed by server. Signature is consistent with the kernel OOM-killer leaving sshd memory-starved (terraform destroy still works, but no result is captured). Reproduced on:

• RediSearch/RediSearch 8.2 / 8.4 aarch64
• redislabsdev/RediSearch v8.2.9 / v8.4.5 aarch64 + v2.8.35 / v2.10.28 x86_64
• ftsb-10K (read-only, 10 k docs) on the same setups runs cleanly → workload-specific.

Setting maxmemory ahead of time means Redis returns OOM command not allowed when used memory > 'maxmemory' instead of letting the kernel decide. Run completes with a partial-result warning, instance stays reachable, comparison data still flows to RTS.

Cap formula

cap = min(pct% * total_system_memory, total - 4 GiB)   if total >  4 GiB
    = pct% * total_system_memory                       if total <= 4 GiB

The 4 GiB OS-reserve floor only kicks in when the instance has enough memory to support it. Tiny test instances (e.g. CI scratch boxes) keep the simple pct cap.

total_system_memory is read from INFO memory so we don't need an instance-type catalog or /proc/meminfo SSH; it works on every supported Redis version. Failure to read it (older Redis, mocked test env) is logged and skipped — never blocks the run.

If Redis already has a non-zero maxmemory configured, the new cap overwrites it with a WARNING log line. Spec authors who want to set their own absolute byte value in init_commands can do so — the explicit CONFIG SET maxmemory <bytes> they emit later will take precedence over our preface.

What changes

• redisbench_admin/run/args.py: new --maxmemory-pct N (int, default None) and --no-maxmemory-cap (store_true) on common_run_args.
• redisbench_admin/run/common.py: new helpers:
  • _get_maxmemory_pct_from_dbconfig — handles both dict and list dbconfig formats
  • _resolve_maxmemory_pct — CLI > spec precedence
  • _compute_maxmemory_cap_bytes — cap arithmetic + floor
  • apply_maxmemory_cap — orchestrator, called as the first thing inside execute_init_commands
  • execute_init_commands and run_redis_pre_steps gain an optional args=None kwarg (default preserves existing behaviour for any caller that doesn't update).
• redisbench_admin/run_local/local_db.py: passes args through the two run_redis_pre_steps callsites (it already had args from local_db_spin).
• run_remote/remote_db.py: not modified in this PR. The spec-field path still fires (it doesn't need args); CLI override on the remote path is a small follow-up.

Compatibility

• RediSearchEnterprise (disk project): zero behaviour change unless their specs add maxmemory_pct or their pipelines pass --maxmemory-pct.
• Existing RediSearch specs: zero behaviour change.
• Existing local-bench runs without args wiring: zero behaviour change.

Tests

13 new unit tests in tests/test_common.py, all pass:

tests/test_common.py::test_get_maxmemory_pct_from_dbconfig_dict           PASSED
tests/test_common.py::test_get_maxmemory_pct_from_dbconfig_list           PASSED
tests/test_common.py::test_resolve_maxmemory_pct_precedence               PASSED
tests/test_common.py::test_compute_maxmemory_cap_bytes_floor_kicks_in_on_large_instances    PASSED
tests/test_common.py::test_compute_maxmemory_cap_bytes_floor_caps_high_pct_on_smaller_box   PASSED
tests/test_common.py::test_compute_maxmemory_cap_bytes_no_floor_for_tiny_instances          PASSED
tests/test_common.py::test_apply_maxmemory_cap_no_op_without_opt_in       PASSED
tests/test_common.py::test_apply_maxmemory_cap_no_op_when_cli_disables    PASSED
tests/test_common.py::test_apply_maxmemory_cap_sets_from_spec_field       PASSED
tests/test_common.py::test_apply_maxmemory_cap_floor_wins_on_smaller_box  PASSED
tests/test_common.py::test_apply_maxmemory_cap_skips_on_invalid_pct       PASSED
tests/test_common.py::test_apply_maxmemory_cap_skips_when_info_unavailable PASSED
tests/test_common.py::test_apply_maxmemory_cap_overwrites_existing_with_warn   PASSED

Coverage:

• precedence resolution
• both dbconfig formats (dict and list)
• cap arithmetic at 128 GiB (floor doesn't bite), 8 GiB (floor wins), 2 GiB (no floor)
• the 4 no-op paths (no opt-in / --no-maxmemory-cap / invalid pct / INFO memory unavailable)
• the spec-field happy path
• overwrite-with-warn against an existing maxmemory

3 pre-existing tests still need RTS_PORT for a live Redis (test_execute_init_commands, test_common_exporter_logic, test_dbconfig_keyspacelen_check) — confirmed failing identically on origin/master without that env, so unrelated to this change.

Test plan after merge

• Bump tests/benchmarks/requirements.txt in RediSearch / RediSearchEnterprise to a release that includes this change.
• Add maxmemory_pct: 80 to the 4 search-expire-* specs only.
• Re-trigger the expiration runs across the 6 historical tags on redislabsdev/RediSearch (priv PR #241) and confirm the bench-server stays reachable.
• Spot-check 2-3 RediSearchEnterprise specs via run-local to confirm no CONFIG SET maxmemory is emitted when the field is absent.

[JoanFM]

I am not sure is good to have negative logic. It is hard to keep it backwards compatible?

[JoanFM]

if setting None to the maxmemory-pct should be the way to remove it, and it should be the default?