diff --git a/.github/workflows/CI.yml b/.github/workflows/CI.yml
index 4194fc2d..550a8832 100644
--- a/.github/workflows/CI.yml
+++ b/.github/workflows/CI.yml
@@ -12,7 +12,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        rust: [stable, 1.39.0]
+        rust: [stable, 1.40.0]
     steps:
     - uses: actions/checkout@master
     - uses: actions-rs/toolchain@v1
@@ -77,7 +77,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        rust: [stable, beta, nightly, 1.39.0]
+        rust: [stable, beta, nightly, 1.40.0]
     steps:
     - uses: actions/checkout@master
     - uses: actions-rs/toolchain@v1
diff --git a/tower/Cargo.toml b/tower/Cargo.toml
index f15eb14e..a3166296 100644
--- a/tower/Cargo.toml
+++ b/tower/Cargo.toml
@@ -32,7 +32,7 @@ discover = []
 filter = []
 hedge = ["filter", "futures-util", "hdrhistogram", "tokio/time"]
 limit = ["tokio/time"]
-load = ["discover", "tokio/time"]
+load = ["tokio/time"]
 load-shed = []
 make = ["tokio/io-std"]
 ready-cache = ["futures-util", "indexmap", "tokio/sync"]
@@ -73,4 +73,4 @@ all-features = true
 rustdoc-args = ["--cfg", "docsrs"]
 
 [package.metadata.playground]
-features = ["full"]
\ No newline at end of file
+features = ["full"]
diff --git a/tower-balance/examples/demo.rs b/tower/examples/tower-balance.rs
similarity index 88%
rename from tower-balance/examples/demo.rs
rename to tower/examples/tower-balance.rs
index cc22bd73..eb891c0f 100644
--- a/tower-balance/examples/demo.rs
+++ b/tower/examples/tower-balance.rs
@@ -1,21 +1,22 @@
 //! Exercises load balancers with mocked services.
 
-use futures_core::TryStream;
+use futures_core::{Stream, TryStream};
 use futures_util::{stream, stream::StreamExt, stream::TryStreamExt};
 use hdrhistogram::Histogram;
 use pin_project::pin_project;
 use rand::{self, Rng};
+use std::hash::Hash;
 use std::time::Duration;
 use std::{
     pin::Pin,
     task::{Context, Poll},
 };
 use tokio::time::{self, Instant};
+use tower::balance as lb;
+use tower::discover::{Change, Discover};
+use tower::limit::concurrency::ConcurrencyLimit;
+use tower::load;
 use tower::util::ServiceExt;
-use tower_balance as lb;
-use tower_discover::{Change, Discover};
-use tower_limit::concurrency::ConcurrencyLimit;
-use tower_load as load;
 use tower_service::Service;
 
 const REQUESTS: usize = 100_000;
@@ -61,13 +62,15 @@ async fn main() {
         d,
         DEFAULT_RTT,
         decay,
-        load::NoInstrument,
+        load::CompleteOnResponse::default(),
     ));
     run("P2C+PeakEWMA...", pe).await;
 
     let d = gen_disco();
-    let ll =
-        lb::p2c::Balance::from_entropy(load::PendingRequestsDiscover::new(d, load::NoInstrument));
+    let ll = lb::p2c::Balance::from_entropy(load::PendingRequestsDiscover::new(
+        d,
+        load::CompleteOnResponse::default(),
+    ));
     run("P2C+LeastLoaded...", ll).await;
 }
 
@@ -78,20 +81,19 @@ type Key = usize;
 #[pin_project]
 struct Disco<S>(Vec<(Key, S)>);
 
-impl<S> Discover for Disco<S>
+impl<S> Stream for Disco<S>
 where
     S: Service<Req, Response = Rsp, Error = Error>,
 {
-    type Key = Key;
-    type Service = S;
-    type Error = Error;
-    fn poll_discover(
-        self: Pin<&mut Self>,
-        _: &mut Context<'_>,
-    ) -> Poll<Result<Change<Self::Key, Self::Service>, Self::Error>> {
+    type Item = Result<Change<Key, S>, Error>;
+
+    fn poll_next(self: Pin<&mut Self>, _: &mut Context<'_>) -> Poll<Option<Self::Item>> {
         match self.project().0.pop() {
-            Some((k, service)) => Poll::Ready(Ok(Change::Insert(k, service))),
-            None => Poll::Pending,
+            Some((k, service)) => Poll::Ready(Some(Ok(Change::Insert(k, service)))),
+            None => {
+                // there may be more later
+                Poll::Pending
+            }
         }
     }
 }
@@ -132,7 +134,7 @@ async fn run<D>(name: &'static str, lb: lb::p2c::Balance<D, Req>)
 where
     D: Discover + Unpin + Send + 'static,
     D::Error: Into<Error>,
-    D::Key: Clone + Send,
+    D::Key: Clone + Send + Hash,
     D::Service: Service<Req, Response = Rsp> + load::Load + Send,
     <D::Service as Service<Req>>::Error: Into<Error>,
     <D::Service as Service<Req>>::Future: Send,
diff --git a/tower/src/load/completion.rs b/tower/src/load/completion.rs
new file mode 100644
index 00000000..d938d7a6
--- /dev/null
+++ b/tower/src/load/completion.rs
@@ -0,0 +1,92 @@
+//! Application-specific request completion semantics.
+
+use futures_core::ready;
+use pin_project::pin_project;
+use std::{
+    future::Future,
+    pin::Pin,
+    task::{Context, Poll},
+};
+
+/// Attaches `H`-typed completion tracker to `V` typed values.
+///
+/// Handles (of type `H`) are intended to be RAII guards that primarily implement `Drop` and update
+/// load metric state as they are dropped. This trait allows implementors to "forward" the handle
+/// to later parts of the request-handling pipeline, so that the handle is only dropped when the
+/// request has truly completed.
+///
+/// This utility allows load metrics to have a protocol-agnostic means to track streams past their
+/// initial response future. For example, if `V` represents an HTTP response type, an
+/// implementation could add `H`-typed handles to each response's extensions to detect when all the
+/// response's extensions have been dropped.
+///
+/// A base `impl<H, V> TrackCompletion<H, V> for CompleteOnResponse` is provided to drop the handle
+/// once the response future is resolved. This is appropriate when a response is discrete and
+/// cannot comprise multiple messages.
+///
+/// In many cases, the `Output` type is simply `V`. However, `TrackCompletion` may alter the type
+/// in order to instrument it appropriately. For example, an HTTP `TrackCompletion` may modify the
+/// body type: so a `TrackCompletion` that takes values of type `http::Response<A>` may output
+/// values of type `http::Response<B>`.
+pub trait TrackCompletion<H, V>: Clone {
+    /// The instrumented value type.
+    type Output;
+
+    /// Attaches a `H`-typed handle to a `V`-typed value.
+    fn track_completion(&self, handle: H, value: V) -> Self::Output;
+}
+
+/// A `TrackCompletion` implementation that considers the request completed when the response
+/// future is resolved.
+#[derive(Clone, Copy, Debug, Default)]
+#[non_exhaustive]
+pub struct CompleteOnResponse;
+
+/// Attaches a `C`-typed completion tracker to the result of an `F`-typed `Future`.
+#[pin_project]
+#[derive(Debug)]
+pub struct TrackCompletionFuture<F, C, H> {
+    #[pin]
+    future: F,
+    handle: Option<H>,
+    completion: C,
+}
+
+// ===== impl InstrumentFuture =====
+
+impl<F, C, H> TrackCompletionFuture<F, C, H> {
+    /// Wraps a future, propagating the tracker into its value if successful.
+    pub fn new(completion: C, handle: H, future: F) -> Self {
+        TrackCompletionFuture {
+            future,
+            completion,
+            handle: Some(handle),
+        }
+    }
+}
+
+impl<F, C, H, T, E> Future for TrackCompletionFuture<F, C, H>
+where
+    F: Future<Output = Result<T, E>>,
+    C: TrackCompletion<H, T>,
+{
+    type Output = Result<C::Output, E>;
+
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        let this = self.project();
+        let rsp = ready!(this.future.poll(cx))?;
+        let h = this.handle.take().expect("handle");
+        Poll::Ready(Ok(this.completion.track_completion(h, rsp)))
+    }
+}
+
+// ===== CompleteOnResponse =====
+
+impl<H, V> TrackCompletion<H, V> for CompleteOnResponse {
+    type Output = V;
+
+    fn track_completion(&self, handle: H, value: V) -> V {
+        drop(handle);
+        value
+    }
+}
diff --git a/tower/src/load/constant.rs b/tower/src/load/constant.rs
index b9ea6ff4..f153689d 100644
--- a/tower/src/load/constant.rs
+++ b/tower/src/load/constant.rs
@@ -1,17 +1,20 @@
-//! A constant `Load` implementation. Primarily useful for testing.
+//! A constant `Load` implementation.
 
+#[cfg(feature = "discover")]
 use crate::discover::{Change, Discover};
+#[cfg(feature = "discover")]
 use futures_core::{ready, Stream};
-use pin_project::pin_project;
-use std::{
-    pin::Pin,
-    task::{Context, Poll},
-};
-use tower_service::Service;
+#[cfg(feature = "discover")]
+use std::pin::Pin;
 
 use super::Load;
+use pin_project::pin_project;
+use std::task::{Context, Poll};
+use tower_service::Service;
 
-/// Wraps a type so that `Load::load` returns a constant value.
+/// Wraps a type so that it implements `Load` and returns a constant load metric.
+///
+/// This load estimator is primarily useful for testing.
 #[pin_project]
 #[derive(Debug)]
 pub struct Constant<T, M> {
@@ -55,6 +58,7 @@ where
 }
 
 /// Proxies `Discover` such that all changes are wrapped with a constant load.
+#[cfg(feature = "discover")]
 impl<D: Discover + Unpin, M: Copy> Stream for Constant<D, M> {
     type Item = Result<Change<D::Key, Constant<D::Service, M>>, D::Error>;
 
diff --git a/tower/src/load/instrument.rs b/tower/src/load/instrument.rs
deleted file mode 100644
index f9a0b9ac..00000000
--- a/tower/src/load/instrument.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-use futures_core::ready;
-use pin_project::pin_project;
-use std::{
-    future::Future,
-    pin::Pin,
-    task::{Context, Poll},
-};
-
-/// Attaches `I`-typed instruments to `V` typed values.
-///
-/// This utility allows load metrics to have a protocol-agnostic means to track streams
-/// past their initial response future. For example, if `V` represents an HTTP response
-/// type, an implementation could add `H`-typed handles to each response's extensions to
-/// detect when the response is dropped.
-///
-/// Handles are intended to be RAII guards that primarily implement `Drop` and update load
-/// metric state as they are dropped.
-///
-/// A base `impl<H, V> Instrument<H, V> for NoInstrument` is provided to drop the handle
-/// immediately. This is appropriate when a response is discrete and cannot comprise
-/// multiple messages.
-///
-/// In many cases, the `Output` type is simply `V`. However, `Instrument` may alter the
-/// type in order to instrument it appropriately. For example, an HTTP Instrument may
-/// modify the body type: so an `Instrument` that takes values of type `http::Response<A>`
-/// may output values of type `http::Response<B>`.
-pub trait Instrument<H, V>: Clone {
-    /// The instrumented value type.
-    type Output;
-
-    /// Attaches an `H`-typed handle to a `V`-typed value.
-    fn instrument(&self, handle: H, value: V) -> Self::Output;
-}
-
-/// A `Instrument` implementation that drops each instrument immediately.
-#[derive(Clone, Copy, Debug)]
-pub struct NoInstrument;
-
-/// Attaches a `I`-typed instruments to the result of an `F`-typed `Future`.
-#[pin_project]
-#[derive(Debug)]
-pub struct InstrumentFuture<F, I, H> {
-    #[pin]
-    future: F,
-    handle: Option<H>,
-    instrument: I,
-}
-
-// ===== impl InstrumentFuture =====
-
-impl<F, I, H> InstrumentFuture<F, I, H> {
-    /// Wraps a future, instrumenting its value if successful.
-    pub fn new(instrument: I, handle: H, future: F) -> Self {
-        InstrumentFuture {
-            future,
-            instrument,
-            handle: Some(handle),
-        }
-    }
-}
-
-impl<F, I, H, T, E> Future for InstrumentFuture<F, I, H>
-where
-    F: Future<Output = Result<T, E>>,
-    I: Instrument<H, T>,
-{
-    type Output = Result<I::Output, E>;
-
-    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
-        let this = self.project();
-        let rsp = ready!(this.future.poll(cx))?;
-        let h = this.handle.take().expect("handle");
-        Poll::Ready(Ok(this.instrument.instrument(h, rsp)))
-    }
-}
-
-// ===== NoInstrument =====
-
-impl<H, V> Instrument<H, V> for NoInstrument {
-    type Output = V;
-
-    fn instrument(&self, handle: H, value: V) -> V {
-        drop(handle);
-        value
-    }
-}
diff --git a/tower/src/load/mod.rs b/tower/src/load/mod.rs
index fdd1f192..7b45ab06 100644
--- a/tower/src/load/mod.rs
+++ b/tower/src/load/mod.rs
@@ -1,22 +1,85 @@
-//! Abstractions and utilties for measuring a service's load.
+//! Service load measurement
+//!
+//! This module provides the [`Load`] trait, which allows measuring how loaded a service is.
+//! It also provides several wrapper types that measure load in different ways:
+//!
+//!  - [`Constant`] — Always returns the same constant load value for a service.
+//!  - [`PendingRequests`] — Measures load by tracking the number of in-flight requests.
+//!  - [`PeakEwma`] — Measures load using a moving average of the peak latency for the service.
+//!
+//! In general, you will want to use one of these when using the types in [`tower::balance`] which
+//! balance services depending on their load. Which load metric to use depends on your exact
+//! use-case, but the ones above should get you quite far!
+//!
+//! When the `discover` feature is enabled, wrapper types for [`tower::discover::Discover`] that
+//! wrap the discovered services with the given load estimator are also provided.
+//!
+//! # When does a request complete?
+//!
+//! For many applications, the request life-cycle is relatively simple: when a service responds to
+//! a request, that request is done, and the system can forget about it. However, for some
+//! applications, the service may respond to the initial request while other parts of the system
+//! are still acting on that request. In such an application, the system load must take these
+//! requests into account as well, or risk the system underestimating its own load.
+//!
+//! To support these use-cases, the load estimators in this module are parameterized by the
+//! [`TrackCompletion`] trait, with [`CompleteOnResponse`] as the default type. The behavior of
+//! `CompleteOnOnResponse` is what you would normally expect for a request-response cycle: when the
+//! response is produced, the request is considered "finished", and load goes down. This can be
+//! overriden by your own user-defined type to track more complex request completion semantics. See
+//! the documentation for [`tower::load::completion`] for more details.
+//!
+//! # Examples
+//!
+//! ```rust
+//! # #[cfg(feature = "util")]
+//! use tower::util::ServiceExt;
+//! # #[cfg(feature = "util")]
+//! use tower::{load::Load, Service};
+//! # #[cfg(feature = "util")]
+//! async fn simple_balance<S1, S2, R>(
+//!     svc1: &mut S1,
+//!     svc2: &mut S2,
+//!     request: R
+//! ) -> Result<S1::Response, S1::Error>
+//! where
+//!     S1: Load + Service<R>,
+//!     S2: Load<Metric = S1::Metric> + Service<R, Response = S1::Response, Error = S1::Error>
+//! {
+//!     if svc1.load() < svc2.load() {
+//!         svc1.ready_and().await?.call(request).await
+//!     } else {
+//!         svc2.ready_and().await?.call(request).await
+//!     }
+//! }
+//! ```
+// TODO: a custom completion example would be good here
 
+pub mod completion;
 mod constant;
-mod instrument;
 pub mod peak_ewma;
 pub mod pending_requests;
 
 pub use self::{
+    completion::{CompleteOnResponse, TrackCompletion},
     constant::Constant,
-    instrument::{Instrument, InstrumentFuture, NoInstrument},
-    peak_ewma::{PeakEwma, PeakEwmaDiscover},
-    pending_requests::{PendingRequests, PendingRequestsDiscover},
+    peak_ewma::PeakEwma,
+    pending_requests::PendingRequests,
 };
 
-/// Exposes a load metric.
+#[cfg(feature = "discover")]
+pub use self::{peak_ewma::PeakEwmaDiscover, pending_requests::PendingRequestsDiscover};
+
+/// Types that implement this trait can give an estimate of how loaded they are.
+///
+/// See the module documentation for more details.
 pub trait Load {
-    /// A comparable load metric. Lesser values are "preferable" to greater values.
+    /// A comparable load metric.
+    ///
+    /// Lesser values indicate that the service is less loaded, and should be preferred for new
+    /// requests over another service with a higher value.
     type Metric: PartialOrd;
 
-    /// Obtains a service's load.
+    /// Estimate the service's current load.
     fn load(&self) -> Self::Metric;
 }
diff --git a/tower/src/load/peak_ewma.rs b/tower/src/load/peak_ewma.rs
index 2369a844..73397a64 100644
--- a/tower/src/load/peak_ewma.rs
+++ b/tower/src/load/peak_ewma.rs
@@ -1,14 +1,17 @@
-//! A `Load` implementation that PeakEWMA on response latency.
+//! A `Load` implementation that measures load using the PeakEWMA response latency.
 
-use super::Load;
-use super::{Instrument, InstrumentFuture, NoInstrument};
+#[cfg(feature = "discover")]
 use crate::discover::{Change, Discover};
+#[cfg(feature = "discover")]
 use futures_core::{ready, Stream};
+#[cfg(feature = "discover")]
 use pin_project::pin_project;
-use std::{
-    pin::Pin,
-    task::{Context, Poll},
-};
+#[cfg(feature = "discover")]
+use std::pin::Pin;
+
+use super::completion::{CompleteOnResponse, TrackCompletion, TrackCompletionFuture};
+use super::Load;
+use std::task::{Context, Poll};
 use std::{
     sync::{Arc, Mutex},
     time::Duration,
@@ -17,7 +20,7 @@ use tokio::time::Instant;
 use tower_service::Service;
 use tracing::trace;
 
-/// Wraps an `S`-typed Service with Peak-EWMA load measurement.
+/// Measures the load of the underlying service using Peak-EWMA load measurement.
 ///
 /// `PeakEwma` implements `Load` with the `Cost` metric that estimates the amount of
 /// pending work to an endpoint. Work is calculated by multiplying the
@@ -26,11 +29,6 @@ use tracing::trace;
 /// worst-case latencies. Over time, the peak latency value decays towards the moving
 /// average of latencies to the endpoint.
 ///
-/// As requests are sent to the underlying service, an `I`-typed instrumentation strategy
-/// is used to track responses to measure latency in an application-specific way. The
-/// default strategy measures latency as the elapsed time from the request being issued to
-/// the underlying service to the response future being satisfied (or dropped).
-///
 /// When no latency information has been measured for an endpoint, an arbitrary default
 /// RTT of 1 second is used to prevent the endpoint from being overloaded before a
 /// meaningful baseline can be established..
@@ -43,22 +41,23 @@ use tracing::trace;
 /// [finagle]:
 /// https://github.com/twitter/finagle/blob/9cc08d15216497bb03a1cafda96b7266cfbbcff1/finagle-core/src/main/scala/com/twitter/finagle/loadbalancer/PeakEwma.scala
 #[derive(Debug)]
-pub struct PeakEwma<S, I = NoInstrument> {
+pub struct PeakEwma<S, C = CompleteOnResponse> {
     service: S,
     decay_ns: f64,
     rtt_estimate: Arc<Mutex<RttEstimate>>,
-    instrument: I,
+    completion: C,
 }
 
-/// Wraps a `D`-typed stream of discovery updates with `PeakEwma`.
+/// Wraps a `D`-typed stream of discovered services with `PeakEwma`.
 #[pin_project]
 #[derive(Debug)]
-pub struct PeakEwmaDiscover<D, I = NoInstrument> {
+#[cfg(feature = "discover")]
+pub struct PeakEwmaDiscover<D, C = CompleteOnResponse> {
     #[pin]
     discover: D,
     decay_ns: f64,
     default_rtt: Duration,
-    instrument: I,
+    completion: C,
 }
 
 /// Represents the relative cost of communicating with a service.
@@ -87,65 +86,14 @@ const NANOS_PER_MILLI: f64 = 1_000_000.0;
 
 // ===== impl PeakEwma =====
 
-impl<D, I> PeakEwmaDiscover<D, I> {
-    /// Wraps a `D`-typed `Discover` so that services have a `PeakEwma` load metric.
-    ///
-    /// The provided `default_rtt` is used as the default RTT estimate for newly
-    /// added services.
-    ///
-    /// They `decay` value determines over what time period a RTT estimate should
-    /// decay.
-    pub fn new<Request>(discover: D, default_rtt: Duration, decay: Duration, instrument: I) -> Self
-    where
-        D: Discover,
-        D::Service: Service<Request>,
-        I: Instrument<Handle, <D::Service as Service<Request>>::Response>,
-    {
-        PeakEwmaDiscover {
-            discover,
-            decay_ns: nanos(decay),
-            default_rtt,
-            instrument,
-        }
-    }
-}
-
-impl<D, I> Stream for PeakEwmaDiscover<D, I>
-where
-    D: Discover,
-    I: Clone,
-{
-    type Item = Result<Change<D::Key, PeakEwma<D::Service, I>>, D::Error>;
-
-    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
-        let this = self.project();
-        let change = match ready!(this.discover.poll_discover(cx)).transpose()? {
-            None => return Poll::Ready(None),
-            Some(Change::Remove(k)) => Change::Remove(k),
-            Some(Change::Insert(k, svc)) => {
-                let peak_ewma = PeakEwma::new(
-                    svc,
-                    *this.default_rtt,
-                    *this.decay_ns,
-                    this.instrument.clone(),
-                );
-                Change::Insert(k, peak_ewma)
-            }
-        };
-
-        Poll::Ready(Some(Ok(change)))
-    }
-}
-
-// ===== impl PeakEwma =====
-
-impl<S, I> PeakEwma<S, I> {
-    fn new(service: S, default_rtt: Duration, decay_ns: f64, instrument: I) -> Self {
+impl<S, C> PeakEwma<S, C> {
+    /// Wraps an `S`-typed service so that its load is tracked by the EWMA of its peak latency.
+    pub fn new(service: S, default_rtt: Duration, decay_ns: f64, completion: C) -> Self {
         Self {
             service,
             decay_ns,
             rtt_estimate: Arc::new(Mutex::new(RttEstimate::new(nanos(default_rtt)))),
-            instrument,
+            completion,
         }
     }
 
@@ -158,29 +106,29 @@ impl<S, I> PeakEwma<S, I> {
     }
 }
 
-impl<S, I, Request> Service<Request> for PeakEwma<S, I>
+impl<S, C, Request> Service<Request> for PeakEwma<S, C>
 where
     S: Service<Request>,
-    I: Instrument<Handle, S::Response>,
+    C: TrackCompletion<Handle, S::Response>,
 {
-    type Response = I::Output;
+    type Response = C::Output;
     type Error = S::Error;
-    type Future = InstrumentFuture<S::Future, I, Handle>;
+    type Future = TrackCompletionFuture<S::Future, C, Handle>;
 
     fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
         self.service.poll_ready(cx)
     }
 
     fn call(&mut self, req: Request) -> Self::Future {
-        InstrumentFuture::new(
-            self.instrument.clone(),
+        TrackCompletionFuture::new(
+            self.completion.clone(),
             self.handle(),
             self.service.call(req),
         )
     }
 }
 
-impl<S, I> Load for PeakEwma<S, I> {
+impl<S, C> Load for PeakEwma<S, C> {
     type Metric = Cost;
 
     fn load(&self) -> Self::Metric {
@@ -201,13 +149,67 @@ impl<S, I> Load for PeakEwma<S, I> {
     }
 }
 
-impl<S, I> PeakEwma<S, I> {
+impl<S, C> PeakEwma<S, C> {
     fn update_estimate(&self) -> f64 {
         let mut rtt = self.rtt_estimate.lock().expect("peak ewma prior_estimate");
         rtt.decay(self.decay_ns)
     }
 }
 
+// ===== impl PeakEwmaDiscover =====
+
+#[cfg(feature = "discover")]
+impl<D, C> PeakEwmaDiscover<D, C> {
+    /// Wraps a `D`-typed `Discover` so that services have a `PeakEwma` load metric.
+    ///
+    /// The provided `default_rtt` is used as the default RTT estimate for newly
+    /// added services.
+    ///
+    /// They `decay` value determines over what time period a RTT estimate should
+    /// decay.
+    pub fn new<Request>(discover: D, default_rtt: Duration, decay: Duration, completion: C) -> Self
+    where
+        D: Discover,
+        D::Service: Service<Request>,
+        C: TrackCompletion<Handle, <D::Service as Service<Request>>::Response>,
+    {
+        PeakEwmaDiscover {
+            discover,
+            decay_ns: nanos(decay),
+            default_rtt,
+            completion,
+        }
+    }
+}
+
+#[cfg(feature = "discover")]
+impl<D, C> Stream for PeakEwmaDiscover<D, C>
+where
+    D: Discover,
+    C: Clone,
+{
+    type Item = Result<Change<D::Key, PeakEwma<D::Service, C>>, D::Error>;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let this = self.project();
+        let change = match ready!(this.discover.poll_discover(cx)).transpose()? {
+            None => return Poll::Ready(None),
+            Some(Change::Remove(k)) => Change::Remove(k),
+            Some(Change::Insert(k, svc)) => {
+                let peak_ewma = PeakEwma::new(
+                    svc,
+                    *this.default_rtt,
+                    *this.decay_ns,
+                    this.completion.clone(),
+                );
+                Change::Insert(k, peak_ewma)
+            }
+        };
+
+        Poll::Ready(Some(Ok(change)))
+    }
+}
+
 // ===== impl RttEstimate =====
 
 impl RttEstimate {
@@ -336,7 +338,7 @@ mod tests {
             Svc,
             Duration::from_millis(10),
             NANOS_PER_MILLI * 1_000.0,
-            NoInstrument,
+            CompleteOnResponse,
         );
         let Cost(load) = svc.load();
         assert_eq!(load, 10.0 * NANOS_PER_MILLI);
@@ -360,7 +362,7 @@ mod tests {
             Svc,
             Duration::from_millis(20),
             NANOS_PER_MILLI * 1_000.0,
-            NoInstrument,
+            CompleteOnResponse,
         );
         assert_eq!(svc.load(), Cost(20.0 * NANOS_PER_MILLI));
 
diff --git a/tower/src/load/pending_requests.rs b/tower/src/load/pending_requests.rs
index a53950c8..ec7b9886 100644
--- a/tower/src/load/pending_requests.rs
+++ b/tower/src/load/pending_requests.rs
@@ -1,37 +1,40 @@
-//! A `Load` implementation that uses the count of in-flight requests.
+//! A `Load` implementation that measures load using the number of in-flight requests.
 
-use super::Load;
-use super::{Instrument, InstrumentFuture, NoInstrument};
+#[cfg(feature = "discover")]
 use crate::discover::{Change, Discover};
+#[cfg(feature = "discover")]
 use futures_core::{ready, Stream};
+#[cfg(feature = "discover")]
 use pin_project::pin_project;
+#[cfg(feature = "discover")]
+use std::pin::Pin;
+
+use super::completion::{CompleteOnResponse, TrackCompletion, TrackCompletionFuture};
+use super::Load;
 use std::sync::Arc;
-use std::{
-    pin::Pin,
-    task::{Context, Poll},
-};
+use std::task::{Context, Poll};
 use tower_service::Service;
 
-/// Expresses load based on the number of currently-pending requests.
+/// Measures the load of the underlying service using the number of currently-pending requests.
 #[derive(Debug)]
-pub struct PendingRequests<S, I = NoInstrument> {
+pub struct PendingRequests<S, C = CompleteOnResponse> {
     service: S,
     ref_count: RefCount,
-    instrument: I,
+    completion: C,
 }
 
-/// Shared between instances of `PendingRequests` and `Handle` to track active
-/// references.
+/// Shared between instances of `PendingRequests` and `Handle` to track active references.
 #[derive(Clone, Debug, Default)]
 struct RefCount(Arc<()>);
 
-/// Wraps `inner`'s services with `PendingRequests`.
+/// Wraps a `D`-typed stream of discovered services with `PendingRequests`.
 #[pin_project]
 #[derive(Debug)]
-pub struct PendingRequestsDiscover<D, I = NoInstrument> {
+#[cfg(feature = "discover")]
+pub struct PendingRequestsDiscover<D, C = CompleteOnResponse> {
     #[pin]
     discover: D,
-    instrument: I,
+    completion: C,
 }
 
 /// Represents the number of currently-pending requests to a given service.
@@ -44,11 +47,12 @@ pub struct Handle(RefCount);
 
 // ===== impl PendingRequests =====
 
-impl<S, I> PendingRequests<S, I> {
-    fn new(service: S, instrument: I) -> Self {
+impl<S, C> PendingRequests<S, C> {
+    /// Wraps an `S`-typed service so that its load is tracked by the number of pending requests.
+    pub fn new(service: S, completion: C) -> Self {
         Self {
             service,
-            instrument,
+            completion,
             ref_count: RefCount::default(),
         }
     }
@@ -58,7 +62,7 @@ impl<S, I> PendingRequests<S, I> {
     }
 }
 
-impl<S, I> Load for PendingRequests<S, I> {
+impl<S, C> Load for PendingRequests<S, C> {
     type Metric = Count;
 
     fn load(&self) -> Count {
@@ -67,22 +71,22 @@ impl<S, I> Load for PendingRequests<S, I> {
     }
 }
 
-impl<S, I, Request> Service<Request> for PendingRequests<S, I>
+impl<S, C, Request> Service<Request> for PendingRequests<S, C>
 where
     S: Service<Request>,
-    I: Instrument<Handle, S::Response>,
+    C: TrackCompletion<Handle, S::Response>,
 {
-    type Response = I::Output;
+    type Response = C::Output;
     type Error = S::Error;
-    type Future = InstrumentFuture<S::Future, I, Handle>;
+    type Future = TrackCompletionFuture<S::Future, C, Handle>;
 
     fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
         self.service.poll_ready(cx)
     }
 
     fn call(&mut self, req: Request) -> Self::Future {
-        InstrumentFuture::new(
-            self.instrument.clone(),
+        TrackCompletionFuture::new(
+            self.completion.clone(),
             self.handle(),
             self.service.call(req),
         )
@@ -91,27 +95,29 @@ where
 
 // ===== impl PendingRequestsDiscover =====
 
-impl<D, I> PendingRequestsDiscover<D, I> {
+#[cfg(feature = "discover")]
+impl<D, C> PendingRequestsDiscover<D, C> {
     /// Wraps a `Discover``, wrapping all of its services with `PendingRequests`.
-    pub fn new<Request>(discover: D, instrument: I) -> Self
+    pub fn new<Request>(discover: D, completion: C) -> Self
     where
         D: Discover,
         D::Service: Service<Request>,
-        I: Instrument<Handle, <D::Service as Service<Request>>::Response>,
+        C: TrackCompletion<Handle, <D::Service as Service<Request>>::Response>,
     {
         Self {
             discover,
-            instrument,
+            completion,
         }
     }
 }
 
-impl<D, I> Stream for PendingRequestsDiscover<D, I>
+#[cfg(feature = "discover")]
+impl<D, C> Stream for PendingRequestsDiscover<D, C>
 where
     D: Discover,
-    I: Clone,
+    C: Clone,
 {
-    type Item = Result<Change<D::Key, PendingRequests<D::Service, I>>, D::Error>;
+    type Item = Result<Change<D::Key, PendingRequests<D::Service, C>>, D::Error>;
 
     /// Yields the next discovery change set.
     fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
@@ -120,7 +126,7 @@ where
         let this = self.project();
         let change = match ready!(this.discover.poll_discover(cx)).transpose()? {
             None => return Poll::Ready(None),
-            Some(Insert(k, svc)) => Insert(k, PendingRequests::new(svc, this.instrument.clone())),
+            Some(Insert(k, svc)) => Insert(k, PendingRequests::new(svc, this.completion.clone())),
             Some(Remove(k)) => Remove(k),
         };
 
@@ -159,7 +165,7 @@ mod tests {
 
     #[test]
     fn default() {
-        let mut svc = PendingRequests::new(Svc, NoInstrument);
+        let mut svc = PendingRequests::new(Svc, CompleteOnResponse);
         assert_eq!(svc.load(), Count(0));
 
         let rsp0 = svc.call(());
@@ -176,12 +182,12 @@ mod tests {
     }
 
     #[test]
-    fn instrumented() {
+    fn with_completion() {
         #[derive(Clone)]
         struct IntoHandle;
-        impl Instrument<Handle, ()> for IntoHandle {
+        impl TrackCompletion<Handle, ()> for IntoHandle {
             type Output = Handle;
-            fn instrument(&self, i: Handle, (): ()) -> Handle {
+            fn track_completion(&self, i: Handle, (): ()) -> Handle {
                 i
             }
         }