refactor(stats): own retention memory estimates

crates/irtt-cli/src/cmd/client/run.rs

@@ -31,8 +31,6 @@ use crate::shared::client::{
 
 use irtt_stats::{StatsCollector, StatsConfig};
 
-const FINITE_STATS_BYTES_PER_PROBE: u64 = 500;
-
 const MIB: u64 = 1024 * 1024;
 
 const GIB: u64 = 1024 * MIB;
@@ -356,9 +354,12 @@ fn finite_stats_memory_warning(args: &ClientArgs, target_count: usize) -> Option
         return None;
     }
     let target_count = u64::try_from(target_count).unwrap_or(u64::MAX);
-    let estimated_bytes = expected_probe_count(args.duration, args.interval)
-        .saturating_mul(FINITE_STATS_BYTES_PER_PROBE)
-        .saturating_mul(target_count);
+    let total_probe_count =
+        expected_probe_count(args.duration, args.interval).saturating_mul(target_count);
+    // Ask the stats crate what the configuration this run will actually use
+    // retains; the CLI owns the probe count and the thresholds, not the
+    // retention model.
+    let estimated_bytes = stats_config(false).estimated_retained_bytes(total_probe_count);
     if estimated_bytes < FINITE_STATS_MEMORY_WARNING_BYTES {
         return None;
     }
@@ -385,11 +386,10 @@ mod tests {
     };
     use std::sync::Arc;
 
-    #[test]
-    fn finite_stats_memory_warning_accounts_for_target_count() {
+    fn long_finite_args() -> ClientArgs {
         use clap::Parser;
 
-        let args = ClientArgs::try_parse_from([
+        ClientArgs::try_parse_from([
             "irtt-cli",
             "--duration",
             "4200s",
@@ -397,33 +397,46 @@ mod tests {
             "100ms",
             "127.0.0.1:2112",
         ])
-        .unwrap();
+        .unwrap()
+    }
+
+    /// Smallest target count whose aggregate estimate reaches the warning
+    /// threshold, derived from the stats crate's own estimate rather than from
+    /// any bytes-per-probe assumption held here.
+    fn first_warning_target_count(args: &ClientArgs) -> usize {
+        let probes = expected_probe_count(args.duration, args.interval);
+        (1..=1024)
+            .find(|targets| {
+                let total = probes.saturating_mul(*targets as u64);
+                stats_config(false).estimated_retained_bytes(total)
+                    >= FINITE_STATS_MEMORY_WARNING_BYTES
+            })
+            .expect("a long finite run should cross the warning threshold within 1024 targets")
+    }
+
+    #[test]
+    fn finite_stats_memory_warning_accounts_for_target_count() {
+        let args = long_finite_args();
+        let crossing = first_warning_target_count(&args);
 
         assert!(
-            finite_stats_memory_warning(&args, 1).is_none(),
-            "a single target's estimate should stay below the warning threshold"
+            crossing > 1,
+            "this run should need more than one target to cross the threshold, \
+             otherwise the test proves nothing about target-count accounting"
         );
         assert!(
-            finite_stats_memory_warning(&args, 7).is_some(),
-            "seven targets' aggregate estimate should cross the warning threshold"
+            finite_stats_memory_warning(&args, crossing - 1).is_none(),
+            "an aggregate estimate below the threshold should not warn"
+        );
+        assert!(
+            finite_stats_memory_warning(&args, crossing).is_some(),
+            "an aggregate estimate at the threshold should warn"
         );
     }
 
     #[test]
     fn finite_stats_memory_warning_saturates_target_count_multiplication() {
-        use clap::Parser;
-
-        let args = ClientArgs::try_parse_from([
-            "irtt-cli",
-            "--duration",
-            "4200s",
-            "--interval",
-            "100ms",
-            "127.0.0.1:2112",
-        ])
-        .unwrap();
-
-        assert!(finite_stats_memory_warning(&args, usize::MAX).is_some());
+        assert!(finite_stats_memory_warning(&long_finite_args(), usize::MAX).is_some());
     }
 
     #[test]

crates/irtt-stats/src/core.rs

@@ -9,7 +9,7 @@ use crate::{
     PacketCounts, RttStats, SampleMode, ServerProcessingStats, Snapshot,
 };
 
-const CONTINUOUS_SEQUENCE_LIMIT: usize = 4096;
+pub(crate) const CONTINUOUS_SEQUENCE_LIMIT: usize = 4096;
 
 #[derive(Debug, Clone, PartialEq)]
 pub(crate) struct CoreStats {

crates/irtt-stats/src/lib.rs

@@ -15,6 +15,7 @@ mod core;
 mod ipdv;
 mod loss;
 mod normalization;
+mod retention;
 mod rolling;
 mod time_stats;
 
@@ -68,6 +69,55 @@ impl StatsConfig {
             late_replies: LateReplyMode::Measure,
         }
     }
+
+    /// Returns an approximate estimate of the bytes a [`StatsCollector`] with
+    /// this configuration retains after processing `probe_count` probes.
+    ///
+    /// This is a planning aid — it exists so callers can warn about a run
+    /// before starting it — and deliberately not allocator accounting. It is
+    /// derived from the structures this crate retains and a fixed
+    /// capacity-headroom factor for the amortized-doubling containers those
+    /// structures use, so it errs toward overestimating.
+    ///
+    /// # What is included
+    ///
+    /// - Exact timing samples, for the metrics that retain them.
+    /// - Adjacent-sequence IPDV tracking state.
+    /// - Count-based rolling storage, when [`StatsConfig::rolling_count`] is
+    ///   set, assuming a probe contributes one send event and one unique
+    ///   reply event.
+    ///
+    /// # What is excluded
+    ///
+    /// - Time-based rolling storage. [`StatsConfig::rolling_time`] retains
+    ///   whatever arrives inside its window, which cannot be bounded from a
+    ///   probe count alone without assuming traffic timing, so it does not
+    ///   contribute to this estimate.
+    /// - Everything outside this crate: sockets, client session state,
+    ///   rendered output, and caller structures.
+    ///
+    /// # Sample mode
+    ///
+    /// [`SampleMode::Exact`] retains one value per metric per measurable
+    /// reply and keeps unbounded IPDV state, so the estimate grows with
+    /// `probe_count`. [`SampleMode::RunningOnly`] retains no exact samples
+    /// and bounds its IPDV state, so the estimate reaches a ceiling and stops
+    /// growing.
+    ///
+    /// # Optional measurements
+    ///
+    /// Adjusted RTT, one-way delay, and send/receive IPDV are only measured
+    /// when the negotiated session supplies the necessary timestamps. The
+    /// estimate assumes all of them are available, which is the upper bound.
+    ///
+    /// # Saturation
+    ///
+    /// The result is deterministic and computed with saturating arithmetic, so
+    /// an enormous `probe_count` saturates at [`u64::MAX`] rather than
+    /// wrapping or panicking.
+    pub fn estimated_retained_bytes(&self, probe_count: u64) -> u64 {
+        retention::estimated_retained_bytes(self, probe_count)
+    }
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]

crates/irtt-stats/src/retention.rs

@@ -0,0 +1,117 @@
+//! Approximate estimates of the storage a statistics configuration retains.
+//!
+//! The estimate is derived from the structures this crate actually keeps, so
+//! retention changes and the planning estimate stay in one place. It models
+//! live element storage plus a fixed capacity-headroom factor; it is not
+//! allocator accounting.
+
+use std::mem::size_of;
+
+use crate::core::CONTINUOUS_SEQUENCE_LIMIT;
+use crate::ipdv::IpdvSample;
+use crate::normalization::{ReplySample, StatsEvent};
+use crate::{SampleMode, StatsConfig};
+
+/// Capacity headroom applied to the live element bytes.
+///
+/// `Vec`, `VecDeque`, `HashMap`, and `HashSet` all grow by doubling, so a
+/// container can hold up to roughly twice the bytes its live elements need.
+/// Assuming the worst case keeps the estimate conservative rather than
+/// falsely precise.
+const GROWTH_FACTOR: u64 = 2;
+
+/// Approximate per-entry control overhead for the hashed containers.
+const HASH_ENTRY_OVERHEAD: u64 = 1;
+
+/// Exact-sample vectors that retain one value per measurable unique reply.
+///
+/// In [`SampleMode::Exact`] these are the primary, raw, and adjusted RTT
+/// metrics, the three IPDV metrics, and the two one-way delay metrics.
+/// `send_call`, `timer_error`, and `server_processing` do not retain exact
+/// samples in any mode.
+///
+/// The adjusted-RTT, one-way, and send/receive IPDV metrics only receive a
+/// sample when the negotiated session supplies the corresponding optional
+/// measurement, so counting all eight is the upper bound.
+const EXACT_SAMPLE_VECS_PER_REPLY: u64 = 8;
+
+/// Normalized events a probe usually contributes to a rolling window: one send
+/// event and one unique reply event.
+const ROLLING_EVENTS_PER_PROBE: u64 = 2;
+
+/// Returns the approximate bytes retained after `probe_count` probes.
+pub(crate) fn estimated_retained_bytes(config: &StatsConfig, probe_count: u64) -> u64 {
+    let cumulative = match config.samples {
+        SampleMode::Exact => exact_bytes(probe_count),
+        SampleMode::RunningOnly => running_only_bytes(probe_count),
+    };
+
+    cumulative
+        .saturating_add(rolling_count_bytes(config, probe_count))
+        .saturating_mul(GROWTH_FACTOR)
+}
+
+/// Exact mode retains one sample per metric and one IPDV tracker entry for
+/// every measurable reply, so its storage grows with the probe count.
+fn exact_bytes(probe_count: u64) -> u64 {
+    let per_reply = EXACT_SAMPLE_VECS_PER_REPLY
+        .saturating_mul(size_of::<i128>() as u64)
+        .saturating_add(ipdv_tracker_bytes_per_sample());
+    probe_count.saturating_mul(per_reply)
+}
+
+/// Running-only mode retains no exact samples and bounds the IPDV tracker at a
+/// fixed number of sequences, so its storage stops growing once that bound is
+/// reached.
+fn running_only_bytes(probe_count: u64) -> u64 {
+    probe_count
+        .min(CONTINUOUS_SEQUENCE_LIMIT as u64)
+        .saturating_mul(ipdv_tracker_bytes_per_sample())
+}
+
+/// The IPDV tracker keys a sample map by sequence and keeps a sequence order
+/// queue and a completed-pair set alongside it.
+fn ipdv_tracker_bytes_per_sample() -> u64 {
+    let sample_entry = (size_of::<u32>() + size_of::<IpdvSample>()) as u64 + HASH_ENTRY_OVERHEAD;
+    let order_entry = size_of::<u32>() as u64;
+    let completed_entry = size_of::<u32>() as u64 + HASH_ENTRY_OVERHEAD;
+    sample_entry
+        .saturating_add(order_entry)
+        .saturating_add(completed_entry)
+}
+
+/// Count-based rolling windows retain whole normalized events, bounded by the
+/// configured event count.
+fn rolling_count_bytes(config: &StatsConfig, probe_count: u64) -> u64 {
+    let Some(limit) = config.rolling_count else {
+        return 0;
+    };
+    let retained = probe_count
+        .saturating_mul(ROLLING_EVENTS_PER_PROBE)
+        .min(limit as u64);
+    retained.saturating_mul(rolling_bytes_per_event())
+}
+
+/// A retained event is the enum itself plus, for a unique reply, the boxed
+/// reply sample it owns.
+fn rolling_bytes_per_event() -> u64 {
+    (size_of::<StatsEvent>() + size_of::<ReplySample>()) as u64
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn per_sample_components_are_non_zero() {
+        assert!(ipdv_tracker_bytes_per_sample() > 0);
+        assert!(rolling_bytes_per_event() > 0);
+    }
+
+    #[test]
+    fn running_only_stops_growing_at_the_bounded_sequence_limit() {
+        let limit = CONTINUOUS_SEQUENCE_LIMIT as u64;
+        assert_eq!(running_only_bytes(limit), running_only_bytes(limit * 100));
+        assert!(running_only_bytes(limit - 1) < running_only_bytes(limit));
+    }
+}

crates/irtt-stats/tests/retention.rs

@@ -0,0 +1,134 @@
+//! Public behavior of the retained-memory estimate.
+
+use std::time::Duration;
+
+use irtt_stats::{SampleMode, StatsConfig};
+
+#[test]
+fn zero_probes_retain_nothing() {
+    assert_eq!(StatsConfig::finite().estimated_retained_bytes(0), 0);
+    assert_eq!(StatsConfig::continuous().estimated_retained_bytes(0), 0);
+}
+
+#[test]
+fn exact_mode_grows_with_probe_count() {
+    let config = StatsConfig::finite();
+    assert_eq!(config.samples, SampleMode::Exact);
+
+    let mut previous = config.estimated_retained_bytes(0);
+    for probes in [1_u64, 1_000, 100_000, 10_000_000] {
+        let current = config.estimated_retained_bytes(probes);
+        assert!(
+            current > previous,
+            "estimate should grow with probe count, but {probes} probes gave {current} \
+             after {previous}"
+        );
+        previous = current;
+    }
+}
+
+#[test]
+fn exact_mode_scales_linearly_with_probe_count() {
+    let config = StatsConfig::finite();
+    let single = config.estimated_retained_bytes(1);
+
+    assert_eq!(config.estimated_retained_bytes(1_000), single * 1_000);
+}
+
+#[test]
+fn enormous_probe_counts_saturate_instead_of_wrapping() {
+    for config in [StatsConfig::finite(), StatsConfig::continuous()] {
+        // Reaching this at all proves no panic; a wrapping multiplication
+        // would land near zero rather than near the maximum.
+        assert!(
+            config.estimated_retained_bytes(u64::MAX) > 0,
+            "a saturating estimate should not wrap to zero"
+        );
+    }
+
+    assert_eq!(
+        StatsConfig::finite().estimated_retained_bytes(u64::MAX),
+        u64::MAX,
+        "an unbounded exact-mode estimate should saturate at the maximum"
+    );
+}
+
+#[test]
+fn running_only_does_not_report_unbounded_per_probe_retention() {
+    let continuous = StatsConfig::continuous();
+    assert_eq!(continuous.samples, SampleMode::RunningOnly);
+
+    let huge = continuous.estimated_retained_bytes(u64::MAX);
+    let modest = continuous.estimated_retained_bytes(1_000_000);
+
+    assert_eq!(
+        huge, modest,
+        "running-only retention is bounded, so a far larger run should not raise the estimate"
+    );
+    assert!(
+        huge < StatsConfig::finite().estimated_retained_bytes(1_000_000),
+        "running-only should estimate far less than exact mode for the same run"
+    );
+}
+
+#[test]
+fn finite_config_estimate_is_deterministic_and_usable_for_planning() {
+    let config = StatsConfig::finite();
+    let probes = 1_000_000;
+    let first = config.estimated_retained_bytes(probes);
+
+    assert_eq!(first, config.estimated_retained_bytes(probes));
+
+    // A million probes should land in a range a caller can reason about: well
+    // over a megabyte, and well under a terabyte. The exact figure is free to
+    // move with the retention model.
+    assert!(
+        first > 1024 * 1024,
+        "a million probes should estimate more than a MiB, got {first}"
+    );
+    assert!(
+        first < 1024 * 1024 * 1024 * 1024,
+        "a million probes should estimate less than a TiB, got {first}"
+    );
+}
+
+#[test]
+fn rolling_count_adds_bounded_retention() {
+    let probes = 1_000_000;
+    let plain = StatsConfig::finite();
+    let rolling = StatsConfig {
+        rolling_count: Some(10_000),
+        ..plain
+    };
+
+    let with_rolling = rolling.estimated_retained_bytes(probes);
+    assert!(
+        with_rolling > plain.estimated_retained_bytes(probes),
+        "a count-based rolling window retains events the plain config does not"
+    );
+
+    // The window is capped, so a larger run adds nothing further to its share.
+    let rolling_share = with_rolling - plain.estimated_retained_bytes(probes);
+    let larger = 10 * probes;
+    let larger_share =
+        rolling.estimated_retained_bytes(larger) - plain.estimated_retained_bytes(larger);
+    assert_eq!(
+        rolling_share, larger_share,
+        "rolling retention should stop growing once the configured count is reached"
+    );
+}
+
+#[test]
+fn time_based_rolling_is_documented_as_excluded() {
+    let plain = StatsConfig::finite();
+    let timed = StatsConfig {
+        rolling_time: Some(Duration::from_secs(10)),
+        ..plain
+    };
+
+    assert_eq!(
+        timed.estimated_retained_bytes(1_000_000),
+        plain.estimated_retained_bytes(1_000_000),
+        "time-based rolling cannot be bounded from a probe count and is excluded"
+    );
+}