diff --git a/osbase/event/BUILD.bazel b/osbase/event/BUILD.bazel
new file mode 100644
index 0000000..a8a2a0c
--- /dev/null
+++ b/osbase/event/BUILD.bazel
@@ -0,0 +1,9 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "event",
+    srcs = ["event.go"],
+    importpath = "source.monogon.dev/osbase/event",
+    visibility = ["//visibility:public"],
+    deps = ["//osbase/supervisor"],
+)
diff --git a/osbase/event/etcd/BUILD.bazel b/osbase/event/etcd/BUILD.bazel
new file mode 100644
index 0000000..925eeaa
--- /dev/null
+++ b/osbase/event/etcd/BUILD.bazel
@@ -0,0 +1,30 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "etcd",
+    srcs = ["etcd.go"],
+    importpath = "source.monogon.dev/osbase/event/etcd",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//osbase/event",
+        "@com_github_cenkalti_backoff_v4//:backoff",
+        "@io_etcd_go_etcd_client_v3//:client",
+    ],
+)
+
+go_test(
+    name = "etcd_test",
+    srcs = ["etcd_test.go"],
+    embed = [":etcd"],
+    deps = [
+        "//osbase/event",
+        "//osbase/logtree",
+        "@io_etcd_go_etcd_api_v3//v3rpc/rpctypes",
+        "@io_etcd_go_etcd_client_pkg_v3//testutil",
+        "@io_etcd_go_etcd_client_v3//:client",
+        "@io_etcd_go_etcd_tests_v3//integration",
+        "@org_golang_google_grpc//codes",
+        "@org_golang_google_grpc//grpclog",
+        "@org_uber_go_zap//:zap",
+    ],
+)
diff --git a/osbase/event/etcd/etcd.go b/osbase/event/etcd/etcd.go
new file mode 100644
index 0000000..8cf2440
--- /dev/null
+++ b/osbase/event/etcd/etcd.go
@@ -0,0 +1,444 @@
+package etcd
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"sync"
+
+	"github.com/cenkalti/backoff/v4"
+	clientv3 "go.etcd.io/etcd/client/v3"
+
+	"source.monogon.dev/osbase/event"
+)
+
+var (
+	// Type assert that *Value implements event.ValueWatcher. We do this
+	// artificially, as there currently is no code path that needs this to be
+	// strictly true.  However, users of this library might want to rely on the
+	// Value type instead of particular Value implementations.
+	_ event.ValueWatch[StringAt] = &Value[StringAt]{}
+)
+
+// ThinClient is a small wrapper interface to combine
+// clientv3.KV and clientv3.Watcher.
+type ThinClient interface {
+	clientv3.KV
+	clientv3.Watcher
+}
+
+// Value is an 'Event Value' backed in an etcd cluster, accessed over an
+// etcd client. This is a stateless handle and can be copied and shared across
+// goroutines.
+type Value[T any] struct {
+	decoder func(key, value []byte) (T, error)
+	etcd    ThinClient
+	key     string
+	keyEnd  string
+}
+
+type Option struct {
+	rangeEnd string
+}
+
+// Range creates a Value that is backed a range of etcd key/value pairs from
+// 'key' passed to NewValue to 'end' passed to Range.
+//
+// The key range semantics (ie. lexicographic ordering) are the same as in etcd
+// ranges, so for example to retrieve all keys prefixed by `foo/` key should be
+// `foo/` and end should be `foo0`.
+//
+// For any update in the given range, the decoder will be called and its result
+// will trigger the return of a Get() call. The decoder should return a type
+// that lets the user distinguish which of the multiple objects in the range got
+// updated, as the Get() call returns no additional information about the
+// location of the retrieved object by itself.
+//
+// The order of values retrieved by Get() is currently fully arbitrary and must
+// not be relied on. It's possible that in the future the order of updates and
+// the blocking behaviour of Get will be formalized, but this is not yet the
+// case. Instead, the data returned should be treated as eventually consistent
+// with the etcd state.
+//
+// For some uses, it might be necessary to first retrieve all the objects
+// contained within the range before starting to block on updates - in this
+// case, the BacklogOnly option should be used when calling Get.
+func Range(end string) *Option {
+	return &Option{
+		rangeEnd: end,
+	}
+}
+
+// NewValue creates a new Value for a given key(s) in an etcd client. The
+// given decoder will be used to convert bytes retrieved from etcd into the
+// interface{} value retrieved by Get by this value's watcher.
+func NewValue[T any](etcd ThinClient, key string, decoder func(key, value []byte) (T, error), options ...*Option) *Value[T] {
+	res := &Value[T]{
+		decoder: decoder,
+		etcd:    etcd,
+		key:     key,
+		keyEnd:  key,
+	}
+
+	for _, opt := range options {
+		if end := opt.rangeEnd; end != "" {
+			res.keyEnd = end
+		}
+	}
+
+	return res
+}
+
+func DecoderNoop(_, value []byte) ([]byte, error) {
+	return value, nil
+}
+
+func DecoderStringAt(key, value []byte) (StringAt, error) {
+	return StringAt{
+		Key:   string(key),
+		Value: string(value),
+	}, nil
+}
+
+type StringAt struct {
+	Key   string
+	Value string
+}
+
+func (e *Value[T]) Watch() event.Watcher[T] {
+	ctx, ctxC := context.WithCancel(context.Background())
+	return &watcher[T]{
+		Value: *e,
+
+		ctx:  ctx,
+		ctxC: ctxC,
+
+		current: make(map[string][]byte),
+
+		getSem: make(chan struct{}, 1),
+	}
+}
+
+type watcher[T any] struct {
+	// Value copy, used to configure the behaviour of this watcher.
+	Value[T]
+
+	// ctx is the context that expresses the liveness of this watcher. It is
+	// canceled when the watcher is closed, and the etcd Watch hangs off of it.
+	ctx  context.Context
+	ctxC context.CancelFunc
+
+	// getSem is a semaphore used to limit concurrent Get calls and throw an
+	// error if concurrent access is attempted.
+	getSem chan struct{}
+
+	// backlogged is a list of keys retrieved from etcd but not yet returned via
+	// Get. These items are not a replay of all the updates from etcd, but are
+	// already compacted to deduplicate updates to the same object (ie., if the
+	// update stream from etcd is for keys A, B, and A, the backlogged list will
+	// only contain one update for A and B each, with the first update for A being
+	// discarded upon arrival of the second update).
+	//
+	// The keys are an index into the current map, which contains the values
+	// retrieved, including ones that have already been returned via Get. This
+	// persistence allows us to deduplicate spurious updates to the user, in which
+	// etcd returned a new revision of a key, but the data stayed the same.
+	backlogged [][]byte
+	// current map, keyed from etcd key into etcd value at said key. This map
+	// persists alongside an etcd connection, permitting deduplication of spurious
+	// etcd updates even across multiple Get calls.
+	current map[string][]byte
+
+	// prev is the etcd store revision of a previously completed etcd Get/Watch
+	// call, used to resume a Watch call in case of failures.
+	prev *int64
+	// wc is the etcd watch channel, or nil if no channel is yet open.
+	wc clientv3.WatchChan
+
+	// testRaceWG is an optional WaitGroup that, if set, will be waited upon
+	// after the initial KV value retrieval, but before the watch is created.
+	// This is only used for testing.
+	testRaceWG *sync.WaitGroup
+	// testSetupWG is an optional WaitGroup that, if set, will be waited upon
+	// after the etcd watch is created.
+	// This is only used for testing.
+	testSetupWG *sync.WaitGroup
+}
+
+// setup initiates wc (the watch channel from etcd) after retrieving the initial
+// value(s) with a get operation.
+func (w *watcher[T]) setup(ctx context.Context) error {
+	if w.wc != nil {
+		return nil
+	}
+	ranged := w.key != w.keyEnd
+
+	// First, check if some data under this key/range already exists.
+
+	// We use an exponential backoff and retry here as the initial Get can fail
+	// if the cluster is unstable (eg. failing over). We only fail the retry if
+	// the context expires.
+	bo := backoff.NewExponentialBackOff()
+	bo.MaxElapsedTime = 0
+
+	err := backoff.Retry(func() error {
+
+		var getOpts []clientv3.OpOption
+		if ranged {
+			getOpts = append(getOpts, clientv3.WithRange(w.keyEnd))
+		}
+		get, err := w.etcd.Get(ctx, w.key, getOpts...)
+		if err != nil {
+			return fmt.Errorf("when retrieving initial value: %w", err)
+		}
+
+		// Assert that the etcd API is behaving as expected.
+		if !ranged && len(get.Kvs) > 1 {
+			panic("More than one key returned in unary GET response")
+		}
+
+		// After a successful Get, save the revision to watch from and re-build the
+		// backlog from scratch based on what was available in the etcd store at that
+		// time.
+		w.prev = &get.Header.Revision
+
+		w.backlogged = nil
+		w.current = make(map[string][]byte)
+		for _, kv := range get.Kvs {
+			w.backlogged = append(w.backlogged, kv.Key)
+			w.current[string(kv.Key)] = kv.Value
+		}
+		return nil
+
+	}, backoff.WithContext(bo, ctx))
+
+	if w.testRaceWG != nil {
+		w.testRaceWG.Wait()
+	}
+	if err != nil {
+		return err
+	}
+
+	watchOpts := []clientv3.OpOption{
+		clientv3.WithRev(*w.prev + 1),
+	}
+	if ranged {
+		watchOpts = append(watchOpts, clientv3.WithRange(w.keyEnd))
+	}
+	w.wc = w.etcd.Watch(w.ctx, w.key, watchOpts...)
+
+	if w.testSetupWG != nil {
+		w.testSetupWG.Wait()
+	}
+	return nil
+}
+
+// backfill blocks until a backlog of items is available. An error is returned
+// if the context is canceled.
+func (w *watcher[T]) backfill(ctx context.Context) error {
+	// Keep watching for watch events.
+	for {
+		var resp *clientv3.WatchResponse
+		select {
+		case r := <-w.wc:
+			resp = &r
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+
+		if resp.Canceled {
+			// Only allow for watches to be canceled due to context
+			// cancellations. Any other error is something we need to handle,
+			// eg. a client close or compaction error.
+			if errors.Is(resp.Err(), ctx.Err()) {
+				return fmt.Errorf("watch canceled: %w", resp.Err())
+			}
+
+			// Attempt to reconnect.
+			if w.wc != nil {
+				// If a wc already exists, close it. This forces a reconnection
+				// by the next setup call.
+				w.ctxC()
+				w.ctx, w.ctxC = context.WithCancel(context.Background())
+				w.wc = nil
+			}
+			if err := w.setup(ctx); err != nil {
+				return fmt.Errorf("failed to setup watcher: %w", err)
+			}
+			continue
+		}
+
+		w.prev = &resp.Header.Revision
+		// Spurious watch event with no update? Keep trying.
+		if len(resp.Events) == 0 {
+			continue
+		}
+
+		// Process updates into compacted list, transforming deletions into value: nil
+		// keyValues. This maps an etcd key into a pointer in the already existing
+		// backlog list. It will then be used to compact all updates into the smallest
+		// backlog possible (by overriding previously backlogged items for a key if this
+		// key is encountered again).
+		//
+		// TODO(q3k): this could be stored in the watcher state to not waste time on
+		// each update, but it's good enough for now.
+
+		// Prepare a set of keys that already exist in the backlog. This will be used
+		// to make sure we don't duplicate backlog entries while maintaining a stable
+		// backlog order.
+		seen := make(map[string]bool)
+		for _, k := range w.backlogged {
+			seen[string(k)] = true
+		}
+
+		for _, ev := range resp.Events {
+			var value []byte
+			switch ev.Type {
+			case clientv3.EventTypeDelete:
+			case clientv3.EventTypePut:
+				value = ev.Kv.Value
+			default:
+				return fmt.Errorf("invalid event type %v", ev.Type)
+			}
+
+			keyS := string(ev.Kv.Key)
+			prev := w.current[keyS]
+			// Short-circuit and skip updates with the same content as already present.
+			// These are sometimes emitted by etcd.
+			if bytes.Equal(prev, value) {
+				continue
+			}
+
+			// Only insert to backlog if not yet present, but maintain order.
+			if !seen[string(ev.Kv.Key)] {
+				w.backlogged = append(w.backlogged, ev.Kv.Key)
+				seen[string(ev.Kv.Key)] = true
+			}
+			// Regardless of backlog list, always update the key to its newest value.
+			w.current[keyS] = value
+		}
+
+		// Still nothing in backlog? Keep trying.
+		if len(w.backlogged) == 0 {
+			continue
+		}
+
+		return nil
+	}
+}
+
+type GetOption struct {
+	backlogOnly bool
+}
+
+// Get implements the Get method of the Watcher interface.
+// It can return an error in three cases:
+//   - the given context is canceled (in which case, the given error will wrap
+//     the context error)
+//   - the watcher's BytesDecoder returned an error (in which case the error
+//     returned by the BytesDecoder will be returned verbatim)
+//   - it has been called with BacklogOnly and the Watcher has no more local
+//     event data to return (see BacklogOnly for more information on the
+//     semantics of this mode of operation)
+//
+// Note that transient and permanent etcd errors are never returned, and the
+// Get call will attempt to recover from these errors as much as possible. This
+// also means that the user of the Watcher will not be notified if the
+// underlying etcd client disconnects from the cluster, or if the cluster loses
+// quorum.
+//
+// TODO(q3k): implement leases to allow clients to be notified when there are
+// transient cluster/quorum/partition errors, if needed.
+//
+// TODO(q3k): implement internal, limited buffering for backlogged data not yet
+// consumed by client, as etcd client library seems to use an unbound buffer in
+// case this happens ( see: watcherStream.buf in clientv3).
+func (w *watcher[T]) Get(ctx context.Context, opts ...event.GetOption[T]) (T, error) {
+	var empty T
+	select {
+	case w.getSem <- struct{}{}:
+	default:
+		return empty, fmt.Errorf("cannot Get() concurrently on a single waiter")
+	}
+	defer func() {
+		<-w.getSem
+	}()
+
+	backlogOnly := false
+	var predicate func(t T) bool
+	for _, opt := range opts {
+		if opt.Predicate != nil {
+			predicate = opt.Predicate
+		}
+		if opt.BacklogOnly {
+			backlogOnly = true
+		}
+	}
+
+	ranged := w.key != w.keyEnd
+	if ranged && predicate != nil {
+		return empty, errors.New("filtering unimplemented for ranged etcd values")
+	}
+	if backlogOnly && predicate != nil {
+		return empty, errors.New("filtering unimplemented for backlog-only requests")
+	}
+
+	for {
+		v, err := w.getUnlocked(ctx, ranged, backlogOnly)
+		if err != nil {
+			return empty, err
+		}
+		if predicate == nil || predicate(v) {
+			return v, nil
+		}
+	}
+}
+
+func (w *watcher[T]) getUnlocked(ctx context.Context, ranged, backlogOnly bool) (T, error) {
+	var empty T
+	// Early check for context cancelations, preventing spurious contact with etcd
+	// if there's no need to.
+	if w.ctx.Err() != nil {
+		return empty, w.ctx.Err()
+	}
+
+	if err := w.setup(ctx); err != nil {
+		return empty, fmt.Errorf("when setting up watcher: %w", err)
+	}
+
+	if backlogOnly && len(w.backlogged) == 0 {
+		return empty, event.ErrBacklogDone
+	}
+
+	// Update backlog from etcd if needed.
+	if len(w.backlogged) < 1 {
+		err := w.backfill(ctx)
+		if err != nil {
+			return empty, fmt.Errorf("when watching for new value: %w", err)
+		}
+	}
+	// Backlog is now guaranteed to contain at least one element.
+
+	if !ranged {
+		// For non-ranged queries, drain backlog fully.
+		if len(w.backlogged) != 1 {
+			panic(fmt.Sprintf("multiple keys in nonranged value: %v", w.backlogged))
+		}
+		k := w.backlogged[0]
+		v := w.current[string(k)]
+		w.backlogged = nil
+		return w.decoder(k, v)
+	} else {
+		// For ranged queries, pop one ranged query off the backlog.
+		k := w.backlogged[0]
+		v := w.current[string(k)]
+		w.backlogged = w.backlogged[1:]
+		return w.decoder(k, v)
+	}
+}
+
+func (w *watcher[T]) Close() error {
+	w.ctxC()
+	return nil
+}
diff --git a/osbase/event/etcd/etcd_test.go b/osbase/event/etcd/etcd_test.go
new file mode 100644
index 0000000..ee3672f
--- /dev/null
+++ b/osbase/event/etcd/etcd_test.go
@@ -0,0 +1,865 @@
+package etcd
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"log"
+	"os"
+	"strconv"
+	"sync"
+	"testing"
+	"time"
+
+	"go.etcd.io/etcd/api/v3/v3rpc/rpctypes"
+	"go.etcd.io/etcd/client/pkg/v3/testutil"
+	clientv3 "go.etcd.io/etcd/client/v3"
+	"go.etcd.io/etcd/tests/v3/integration"
+	"go.uber.org/zap"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/grpclog"
+
+	"source.monogon.dev/osbase/event"
+	"source.monogon.dev/osbase/logtree"
+)
+
+var (
+	cluster   *integration.ClusterV3
+	endpoints []string
+)
+
+// TestMain brings up a 3 node etcd cluster for tests to use.
+func TestMain(m *testing.M) {
+	// This logtree's data is not output anywhere.
+	lt := logtree.New()
+
+	cfg := integration.ClusterConfig{
+		Size:                 3,
+		GRPCKeepAliveMinTime: time.Millisecond,
+		LoggerBuilder: func(memberName string) *zap.Logger {
+			dn := logtree.DN("etcd." + memberName)
+			return logtree.Zapify(lt.MustLeveledFor(dn), zap.WarnLevel)
+		},
+	}
+	tb, cancel := testutil.NewTestingTBProthesis("curator")
+	defer cancel()
+	flag.Parse()
+	integration.BeforeTestExternal(tb)
+	grpclog.SetLoggerV2(logtree.GRPCify(lt.MustLeveledFor("grpc")))
+	cluster = integration.NewClusterV3(tb, &cfg)
+	endpoints = make([]string, 3)
+	for i := range endpoints {
+		endpoints[i] = cluster.Client(i).Endpoints()[0]
+	}
+
+	v := m.Run()
+	cluster.Terminate(tb)
+	os.Exit(v)
+}
+
+// setRaceWg creates a new WaitGroup and sets the given watcher to wait on this
+// WG after it performs the initial retrieval of a value from etcd, but before
+// it starts the watcher. This is used to test potential race conditions
+// present between these two steps.
+func setRaceWg[T any](w event.Watcher[T]) *sync.WaitGroup {
+	var wg sync.WaitGroup
+	w.(*watcher[T]).testRaceWG = &wg
+	return &wg
+}
+
+// setSetupWg creates a new WaitGroup and sets the given watcher to wait on
+// thie WG after an etcd watch channel is created. This is used in tests to
+// ensure that the watcher is fully created before it is tested.
+func setSetupWg[T any](w event.Watcher[T]) *sync.WaitGroup {
+	var wg sync.WaitGroup
+	w.(*watcher[T]).testSetupWG = &wg
+	return &wg
+}
+
+// testClient is an etcd connection to the test cluster.
+type testClient struct {
+	client     *clientv3.Client
+}
+
+func newTestClient(t *testing.T) *testClient {
+	t.Helper()
+	cli, err := clientv3.New(clientv3.Config{
+		Endpoints:            endpoints,
+		DialTimeout:          1 * time.Second,
+		DialKeepAliveTime:    1 * time.Second,
+		DialKeepAliveTimeout: 1 * time.Second,
+	})
+	if err != nil {
+		t.Fatalf("clientv3.New: %v", err)
+	}
+
+	return &testClient{
+		client:     cli,
+	}
+}
+
+func (d *testClient) close() {
+	d.client.Close()
+}
+
+// setEndpoints configures which endpoints (from {0,1,2}) the testClient is
+// connected to.
+func (d *testClient) setEndpoints(nums ...uint) {
+	var eps []string
+	for _, num := range nums {
+		eps = append(eps, endpoints[num])
+	}
+	d.client.SetEndpoints(eps...)
+}
+
+// put uses the testClient to store key with a given string value in etcd. It
+// contains retry logic that will block until the put is successful.
+func (d *testClient) put(t *testing.T, key, value string) {
+	t.Helper()
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	for {
+		ctxT, ctxC := context.WithTimeout(ctx, 100*time.Millisecond)
+		_, err := d.client.Put(ctxT, key, value)
+		ctxC()
+		if err == nil {
+			return
+		}
+		if errors.Is(err, ctxT.Err()) {
+			log.Printf("Retrying after %v", err)
+			continue
+		}
+		// Retry on etcd unavailability - this will happen in this code as the
+		// etcd cluster repeatedly loses quorum.
+		var eerr rpctypes.EtcdError
+		if errors.As(err, &eerr) && eerr.Code() == codes.Unavailable {
+			log.Printf("Retrying after %v", err)
+			continue
+		}
+		t.Fatalf("Put: %v", err)
+	}
+
+}
+
+// remove uses the testClient to remove the given key from etcd. It contains
+// retry logic that will block until the removal is successful.
+func (d *testClient) remove(t *testing.T, key string) {
+	t.Helper()
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	_, err := d.client.Delete(ctx, key)
+	if err == nil {
+		return
+	}
+	t.Fatalf("Delete: %v", err)
+}
+
+// expect runs a Get on the given Watcher, ensuring the returned value is a
+// given string.
+func expect(t *testing.T, w event.Watcher[StringAt], value string) {
+	t.Helper()
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	got, err := w.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+
+	if got, want := got.Value, value; got != want {
+		t.Errorf("Wanted value %q, got %q", want, got)
+	}
+}
+
+// expectTimeout ensures that the given watcher blocks on a Get call for at
+// least 100 milliseconds. This is used by tests to attempt to verify that the
+// watcher Get is fully blocked, but can cause false positives (eg. when Get
+// blocks for 101 milliseconds). Thus, this function should be used sparingly
+// and in tests that perform other baseline behaviour checks alongside this
+// test.
+func expectTimeout[T any](t *testing.T, w event.Watcher[T]) {
+	t.Helper()
+	ctx, ctxC := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	got, err := w.Get(ctx)
+	ctxC()
+
+	if !errors.Is(err, ctx.Err()) {
+		t.Fatalf("Expected timeout error, got %v, %v", got, err)
+	}
+}
+
+// wait wraps a watcher into a channel of strings, ensuring that the watcher
+// never errors on Get calls and always returns strings.
+func wait(t *testing.T, w event.Watcher[StringAt]) (chan string, func()) {
+	t.Helper()
+	ctx, ctxC := context.WithCancel(context.Background())
+
+	c := make(chan string)
+
+	go func() {
+		for {
+			got, err := w.Get(ctx)
+			if err != nil && errors.Is(err, ctx.Err()) {
+				return
+			}
+			if err != nil {
+				t.Errorf("Get: %v", err)
+				close(c)
+				return
+			}
+			c <- got.Value
+		}
+	}()
+
+	return c, ctxC
+}
+
+// TestSimple exercises the simplest possible interaction with a watched value.
+func TestSimple(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-simple"
+	value := NewValue(tc.client, k, DecoderStringAt)
+	tc.put(t, k, "one")
+
+	watcher := value.Watch()
+	defer watcher.Close()
+	expect(t, watcher, "one")
+
+	tc.put(t, k, "two")
+	expect(t, watcher, "two")
+
+	tc.put(t, k, "three")
+	tc.put(t, k, "four")
+	tc.put(t, k, "five")
+	tc.put(t, k, "six")
+
+	q, cancel := wait(t, watcher)
+	// Test will hang here if the above value does not receive the set "six".
+	log.Printf("a")
+	for el := range q {
+		log.Printf("%q", el)
+		if el == "six" {
+			break
+		}
+	}
+	log.Printf("b")
+	cancel()
+}
+
+// stringAtGet performs a Get from a Watcher, expecting a stringAt and updating
+// the given map with the retrieved value.
+func stringAtGet(ctx context.Context, t *testing.T, w event.Watcher[StringAt], m map[string]string) {
+	t.Helper()
+
+	vr, err := w.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+	m[vr.Key] = vr.Value
+}
+
+// TestSimpleRange exercises the simplest behaviour of a ranged watcher,
+// retrieving updaates via Get in a fully blocking fashion.
+func TestSimpleRange(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	ks := "test-simple-range/"
+	ke := "test-simple-range0"
+	value := NewValue(tc.client, ks, DecoderStringAt, Range(ke))
+	tc.put(t, ks+"a", "one")
+	tc.put(t, ks+"b", "two")
+	tc.put(t, ks+"c", "three")
+	tc.put(t, ks+"b", "four")
+
+	w := value.Watch()
+	defer w.Close()
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	res := make(map[string]string)
+	stringAtGet(ctx, t, w, res)
+	stringAtGet(ctx, t, w, res)
+	stringAtGet(ctx, t, w, res)
+
+	tc.put(t, ks+"a", "five")
+	tc.put(t, ks+"e", "six")
+
+	stringAtGet(ctx, t, w, res)
+	stringAtGet(ctx, t, w, res)
+
+	for _, te := range []struct {
+		k, w string
+	}{
+		{ks + "a", "five"},
+		{ks + "b", "four"},
+		{ks + "c", "three"},
+		{ks + "e", "six"},
+	} {
+		if want, got := te.w, res[te.k]; want != got {
+			t.Errorf("res[%q]: wanted %q, got %q", te.k, want, got)
+		}
+	}
+}
+
+// TestCancel ensures that watchers can resume after being canceled.
+func TestCancel(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-cancel"
+	value := NewValue(tc.client, k, DecoderStringAt)
+	tc.put(t, k, "one")
+
+	watcher := value.Watch()
+	defer watcher.Close()
+	expect(t, watcher, "one")
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	errs := make(chan error, 1)
+	go func() {
+		_, err := watcher.Get(ctx)
+		errs <- err
+	}()
+	ctxC()
+	if want, got := ctx.Err(), <-errs; !errors.Is(got, want) {
+		t.Fatalf("Wanted err %v, got %v", want, got)
+	}
+
+	// Successfully canceled watch, resuming should continue to work.
+	q, cancel := wait(t, watcher)
+	defer cancel()
+
+	tc.put(t, k, "two")
+	if want, got := "two", <-q; want != got {
+		t.Fatalf("Wanted val %q, got %q", want, got)
+	}
+}
+
+// TestCancelOnGet ensures that a context cancellation on an initial Get (which
+// translates to an etcd Get in a backoff loop) doesn't block.
+func TestCancelOnGet(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-cancel-on-get"
+	value := NewValue(tc.client, k, DecoderStringAt)
+	watcher := value.Watch()
+	tc.put(t, k, "one")
+
+	// Cause partition between client endpoint and rest of cluster. Any read/write
+	// operations will now hang.
+	tc.setEndpoints(0)
+	cluster.Members[0].InjectPartition(t, cluster.Members[1], cluster.Members[2])
+	// Let raft timeouts expire so that the leader is aware a partition has occurred
+	// and stops serving data if it is not part of a quorum anymore.
+	//
+	// Otherwise, if Member[0] was the leader, there will be a window of opportunity
+	// during which it will continue to serve read data even though it has been
+	// partitioned off. This is an effect of how etcd handles linearizable reads:
+	// they go through the leader, but do not go through raft.
+	//
+	// The value is the default etcd leader timeout (1s) + some wiggle room.
+	time.Sleep(time.Second + time.Millisecond*100)
+
+	// Perform the initial Get(), which should attempt to retrieve a KV entry from
+	// the etcd service. This should hang. Unfortunately, there's no easy way to do
+	// this without an arbitrary sleep hoping that the client actually gets to the
+	// underlying etcd.Get call. This can cause false positives (eg. false 'pass'
+	// results) in this test.
+	ctx, ctxC := context.WithCancel(context.Background())
+	errs := make(chan error, 1)
+	go func() {
+		_, err := watcher.Get(ctx)
+		errs <- err
+	}()
+	time.Sleep(time.Second)
+
+	// Now that the etcd.Get is hanging, cancel the context.
+	ctxC()
+	// And now unpartition the cluster, resuming reads.
+	cluster.Members[0].RecoverPartition(t, cluster.Members[1], cluster.Members[2])
+
+	// The etcd.Get() call should've returned with a context cancellation.
+	err := <-errs
+	switch {
+	case err == nil:
+		t.Errorf("watcher.Get() returned no error, wanted context error")
+	case errors.Is(err, ctx.Err()):
+		// Okay.
+	default:
+		t.Errorf("watcher.Get() returned %v, wanted context error", err)
+	}
+}
+
+// TestClientReconnect forces a 'reconnection' of an active watcher from a
+// running member to another member, by stopping the original member and
+// explicitly reconnecting the client to other available members.
+//
+// This doe not reflect a situation expected during Metropolis runtime, as we
+// do not expect splits between an etcd client and its connected member
+// (instead, all etcd clients only connect to their local member). However, it
+// is still an important safety test to perform, and it also exercies the
+// equivalent behaviour of an etcd client re-connecting for any other reason.
+func TestClientReconnect(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+	tc.setEndpoints(0)
+
+	k := "test-client-reconnect"
+	value := NewValue(tc.client, k, DecoderStringAt)
+	tc.put(t, k, "one")
+
+	watcher := value.Watch()
+	defer watcher.Close()
+	expect(t, watcher, "one")
+
+	q, cancel := wait(t, watcher)
+	defer cancel()
+
+	cluster.Members[0].Stop(t)
+	defer cluster.Members[0].Restart(t)
+	cluster.WaitLeader(t)
+
+	tc.setEndpoints(1, 2)
+	tc.put(t, k, "two")
+
+	if want, got := "two", <-q; want != got {
+		t.Fatalf("Watcher received incorrect data after client restart, wanted %q, got %q", want, got)
+	}
+}
+
+// TestClientPartition forces a temporary partition of the etcd member while a
+// watcher is running, updates the value from across the partition, and undoes
+// the partition.
+// The partition is expected to be entirely transparent to the watcher.
+func TestClientPartition(t *testing.T) {
+	tcOne := newTestClient(t)
+	defer tcOne.close()
+	tcOne.setEndpoints(0)
+
+	tcRest := newTestClient(t)
+	defer tcRest.close()
+	tcRest.setEndpoints(1, 2)
+
+	k := "test-client-partition"
+	valueOne := NewValue(tcOne.client, k, DecoderStringAt)
+	watcherOne := valueOne.Watch()
+	defer watcherOne.Close()
+	valueRest := NewValue(tcRest.client, k, DecoderStringAt)
+	watcherRest := valueRest.Watch()
+	defer watcherRest.Close()
+
+	tcRest.put(t, k, "a")
+	expect(t, watcherOne, "a")
+	expect(t, watcherRest, "a")
+
+	cluster.Members[0].InjectPartition(t, cluster.Members[1], cluster.Members[2])
+
+	tcRest.put(t, k, "b")
+	expect(t, watcherRest, "b")
+	expectTimeout(t, watcherOne)
+
+	cluster.Members[0].RecoverPartition(t, cluster.Members[1], cluster.Members[2])
+
+	expect(t, watcherOne, "b")
+	tcRest.put(t, k, "c")
+	expect(t, watcherOne, "c")
+	expect(t, watcherRest, "c")
+
+}
+
+// TestEarlyUse exercises the correct behaviour of the value watcher on a value
+// that is not yet set.
+func TestEarlyUse(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-early-use"
+
+	value := NewValue(tc.client, k, DecoderStringAt)
+	watcher := value.Watch()
+	defer watcher.Close()
+
+	wg := setSetupWg(watcher)
+	wg.Add(1)
+	q, cancel := wait(t, watcher)
+	defer cancel()
+
+	wg.Done()
+
+	tc.put(t, k, "one")
+
+	if want, got := "one", <-q; want != got {
+		t.Fatalf("Expected %q, got %q", want, got)
+	}
+}
+
+// TestRemove exercises the basic functionality of handling deleted values.
+func TestRemove(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-remove"
+	tc.put(t, k, "one")
+
+	value := NewValue(tc.client, k, DecoderStringAt)
+	watcher := value.Watch()
+	defer watcher.Close()
+
+	expect(t, watcher, "one")
+	tc.remove(t, k)
+	expect(t, watcher, "")
+}
+
+// TestRemoveRange exercises the behaviour of a Get on a ranged watcher when a
+// value is removed.
+func TestRemoveRange(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	ks := "test-remove-range/"
+	ke := "test-remove-range0"
+	value := NewValue(tc.client, ks, DecoderStringAt, Range(ke))
+	tc.put(t, ks+"a", "one")
+	tc.put(t, ks+"b", "two")
+	tc.put(t, ks+"c", "three")
+	tc.put(t, ks+"b", "four")
+	tc.remove(t, ks+"c")
+
+	w := value.Watch()
+	defer w.Close()
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	res := make(map[string]string)
+	stringAtGet(ctx, t, w, res)
+	stringAtGet(ctx, t, w, res)
+
+	for _, te := range []struct {
+		k, w string
+	}{
+		{ks + "a", "one"},
+		{ks + "b", "four"},
+		{ks + "c", ""},
+	} {
+		if want, got := te.w, res[te.k]; want != got {
+			t.Errorf("res[%q]: wanted %q, got %q", te.k, want, got)
+		}
+	}
+}
+
+// TestEmptyRace forces the watcher to retrieve an empty value from the K/V
+// store at first, and establishing the watch channel after a new value has
+// been stored in the same place.
+func TestEmptyRace(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-remove-race"
+	tc.put(t, k, "one")
+	tc.remove(t, k)
+
+	value := NewValue(tc.client, k, DecoderStringAt)
+	watcher := value.Watch()
+	defer watcher.Close()
+
+	wg := setRaceWg(watcher)
+	wg.Add(1)
+	q, cancel := wait(t, watcher)
+	defer cancel()
+
+	tc.put(t, k, "two")
+	wg.Done()
+
+	if want, got := "two", <-q; want != got {
+		t.Fatalf("Watcher received incorrect data after client restart, wanted %q, got %q", want, got)
+	}
+}
+
+type errOrInt struct {
+	val int64
+	err error
+}
+
+// TestDecoder exercises the BytesDecoder functionality of the watcher, by
+// creating a value with a decoder that only accepts string-encoded integers
+// that are divisible by three. The test then proceeds to put a handful of
+// values into etcd, ensuring that undecodable values correctly return an error
+// on Get, but that the watcher continues to work after the error has been
+// returned.
+func TestDecoder(t *testing.T) {
+	decoderDivisibleByThree := func(_, value []byte) (int64, error) {
+		num, err := strconv.ParseInt(string(value), 10, 64)
+		if err != nil {
+			return 0, fmt.Errorf("not a valid number")
+		}
+		if (num % 3) != 0 {
+			return 0, fmt.Errorf("not divisible by 3")
+		}
+		return num, nil
+	}
+
+	tc := newTestClient(t)
+	defer tc.close()
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	k := "test-decoder"
+	value := NewValue(tc.client, k, decoderDivisibleByThree)
+	watcher := value.Watch()
+	defer watcher.Close()
+	tc.put(t, k, "3")
+	_, err := watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Initial Get: %v", err)
+	}
+
+	// Stream updates into arbitrarily-bounded test channel.
+	queue := make(chan errOrInt, 100)
+	go func() {
+		for {
+			res, err := watcher.Get(ctx)
+			if err != nil && errors.Is(err, ctx.Err()) {
+				return
+			}
+			if err != nil {
+				queue <- errOrInt{
+					err: err,
+				}
+			} else {
+				queue <- errOrInt{
+					val: res,
+				}
+			}
+		}
+	}()
+
+	var wantList []*int64
+	wantError := func(val string) {
+		wantList = append(wantList, nil)
+		tc.put(t, k, val)
+	}
+	wantValue := func(val string, decoded int64) {
+		wantList = append(wantList, &decoded)
+		tc.put(t, k, val)
+	}
+
+	wantError("")
+	wantValue("9", 9)
+	wantError("foo")
+	wantValue("18", 18)
+	wantError("10")
+	wantValue("27", 27)
+	wantValue("36", 36)
+
+	for i, want := range wantList {
+		q := <-queue
+		if want == nil && q.err == nil {
+			t.Fatalf("Case %d: wanted error, got no error and value %d", i, q.val)
+		}
+		if want != nil && (*want) != q.val {
+			t.Fatalf("Case %d: wanted value %d, got error %v and value %d", i, *want, q.err, q.val)
+		}
+	}
+}
+
+// TestBacklog ensures that the watcher can handle a large backlog of changes
+// in etcd that the client didnt' keep up with, and that whatever final state
+// is available to the client when it actually gets to calling Get().
+func TestBacklog(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	k := "test-backlog"
+	value := NewValue(tc.client, k, DecoderStringAt)
+	watcher := value.Watch()
+	defer watcher.Close()
+
+	tc.put(t, k, "initial")
+	expect(t, watcher, "initial")
+
+	for i := 0; i < 1000; i++ {
+		tc.put(t, k, fmt.Sprintf("val-%d", i))
+	}
+
+	ctx, ctxC := context.WithTimeout(context.Background(), time.Second)
+	defer ctxC()
+	for {
+		valB, err := watcher.Get(ctx)
+		if err != nil {
+			t.Fatalf("Get() returned error before expected final value: %v", err)
+		}
+		if valB.Value == "val-999" {
+			break
+		}
+	}
+}
+
+// TestBacklogRange ensures that the ranged etcd watcher can handle a large
+// backlog of changes in etcd that the client didn't keep up with.
+func TestBacklogRange(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+
+	ks := "test-backlog-range/"
+	ke := "test-backlog-range0"
+	value := NewValue(tc.client, ks, DecoderStringAt, Range(ke))
+	w := value.Watch()
+	defer w.Close()
+
+	for i := 0; i < 100; i++ {
+		if i%2 == 0 {
+			tc.put(t, ks+"a", fmt.Sprintf("val-%d", i))
+		} else {
+			tc.put(t, ks+"b", fmt.Sprintf("val-%d", i))
+		}
+	}
+
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	res := make(map[string]string)
+	stringAtGet(ctx, t, w, res)
+	stringAtGet(ctx, t, w, res)
+
+	for _, te := range []struct {
+		k, w string
+	}{
+		{ks + "a", "val-98"},
+		{ks + "b", "val-99"},
+	} {
+		if want, got := te.w, res[te.k]; want != got {
+			t.Errorf("res[%q]: wanted %q, got %q", te.k, want, got)
+		}
+	}
+}
+
+// TestBacklogOnly exercises the BacklogOnly option for non-ranged watchers,
+// which effectively makes any Get operation non-blocking (but also showcases
+// that unless a Get without BacklogOnly is issues, no new data will appear by
+// itself in the watcher - which is an undocumented implementation detail of the
+// option).
+func TestBacklogOnly(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	k := "test-backlog-only"
+	tc.put(t, k, "initial")
+
+	value := NewValue(tc.client, k, DecoderStringAt)
+	watcher := value.Watch()
+	defer watcher.Close()
+
+	d, err := watcher.Get(ctx, event.BacklogOnly[StringAt]())
+	if err != nil {
+		t.Fatalf("First Get failed: %v", err)
+	}
+	if want, got := "initial", d.Value; want != got {
+		t.Fatalf("First Get: wanted value %q, got %q", want, got)
+	}
+
+	// As expected, next call to Get with BacklogOnly fails - there truly is no new
+	// updates to emit.
+	_, err = watcher.Get(ctx, event.BacklogOnly[StringAt]())
+	if want, got := event.ErrBacklogDone, err; !errors.Is(got, want) {
+		t.Fatalf("Second Get: wanted %v, got %v", want, got)
+	}
+
+	// Implementation detail: even though there is a new value ('second'),
+	// BacklogOnly will still return ErrBacklogDone.
+	tc.put(t, k, "second")
+	_, err = watcher.Get(ctx, event.BacklogOnly[StringAt]())
+	if want, got := event.ErrBacklogDone, err; !errors.Is(got, want) {
+		t.Fatalf("Third Get: wanted %v, got %v", want, got)
+	}
+
+	// ... However, a Get  without BacklogOnly will return the new value.
+	d, err = watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Fourth Get failed: %v", err)
+	}
+	if want, got := "second", d.Value; want != got {
+		t.Fatalf("Fourth Get: wanted value %q, got %q", want, got)
+	}
+}
+
+// TestBacklogOnlyRange exercises the BacklogOnly option for ranged watchers,
+// showcasing how it expected to be used for keeping up with the external state
+// of a range by synchronizing to a local map.
+func TestBacklogOnlyRange(t *testing.T) {
+	tc := newTestClient(t)
+	defer tc.close()
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	ks := "test-backlog-only-range/"
+	ke := "test-backlog-only-range0"
+
+	for i := 0; i < 100; i++ {
+		if i%2 == 0 {
+			tc.put(t, ks+"a", fmt.Sprintf("val-%d", i))
+		} else {
+			tc.put(t, ks+"b", fmt.Sprintf("val-%d", i))
+		}
+	}
+
+	value := NewValue(tc.client, ks, DecoderStringAt, Range(ke))
+	w := value.Watch()
+	defer w.Close()
+
+	// Collect results into a map from key to value.
+	res := make(map[string]string)
+
+	// Run first Get - this is the barrier defining what's part of the backlog.
+	g, err := w.Get(ctx, event.BacklogOnly[StringAt]())
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+	res[g.Key] = g.Value
+
+	// These won't be part of the backlog.
+	tc.put(t, ks+"a", "val-100")
+	tc.put(t, ks+"b", "val-101")
+
+	// Retrieve the rest of the backlog until ErrBacklogDone is returned.
+	nUpdates := 1
+	for {
+		g, err := w.Get(ctx, event.BacklogOnly[StringAt]())
+		if errors.Is(err, event.ErrBacklogDone) {
+			break
+		}
+		if err != nil {
+			t.Fatalf("Get: %v", err)
+		}
+		nUpdates += 1
+		res[g.Key] = g.Value
+	}
+
+	// The backlog should've been compacted to just two entries at their newest
+	// state.
+	if want, got := 2, nUpdates; want != got {
+		t.Fatalf("wanted backlog in %d updates, got it in %d", want, got)
+	}
+
+	for _, te := range []struct {
+		k, w string
+	}{
+		{ks + "a", "val-98"},
+		{ks + "b", "val-99"},
+	} {
+		if want, got := te.w, res[te.k]; want != got {
+			t.Errorf("res[%q]: wanted %q, got %q", te.k, want, got)
+		}
+	}
+}
diff --git a/osbase/event/event.go b/osbase/event/event.go
new file mode 100644
index 0000000..e2e0bd8
--- /dev/null
+++ b/osbase/event/event.go
@@ -0,0 +1,237 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package event defines and implements Event Values, a mechanism in which
+// multiple consumers can watch a value for updates in a reliable way.
+//
+// Values currently are kept in memory (see: MemoryValue), but a future
+// implementation might exist for other storage backends, eg. etcd.
+//
+// # Background and intended use
+//
+// The Event Value library is intended to be used within Metropolis'
+// supervisor-based runnables to communicate state changes to other runnables,
+// while permitting both sides to restart if needed. It grew out of multiple
+// codebases reimplementing an ad-hoc observer pattern, and from the
+// realization that implementing all possible edge cases of such patterns is
+// non-trivial and subject to programming errors. As such, it was turned into a
+// self-standing library.
+//
+// Why not just channels?
+//
+// Plain channels have multiple deficiencies for this usecase:
+//   - Strict FIFO behaviour: all values sent to a channel must be received, and
+//     historic and newest data must be treated in the same way. This means that
+//     a consumer of state changes must process all updates to the value as if
+//     they are the newest, and unable to skip rapid updates when a system is
+//     slowly settling due to a cascading state change.
+//   - Implementation overhead: implementing an observer
+//     registration/unregistration pattern is prone to programming bugs,
+//     especially for features like always first sending the current state to a
+//     new observer.
+//   - Strict buffer size: due to their FIFO nature and the possibility of
+//     consumers not receiving actively, channels would have to buffer all
+//     existing updates, requiring some arbitrary best-guess channel buffer
+//     sizing that would still not prevent blocking writes or data loss in a
+//     worst case scenario.
+//
+// Or, in other words: Go channels are a synchronization primitive, not a
+// ready-made solution to this problem. The Event Value implementation in fact
+// extensively uses Go channels within its implementation as a building block.
+//
+// Why not just condition variables (sync.Cond)?
+//
+// Go's condition variable implementation doesn't fully address our needs
+// either:
+//   - No context/canceling support: once a condition is being Wait()ed on,
+//     this cannot be interrupted. This is especially painful and unwieldy when
+//     dealing with context-heavy code, such as Metropolis.
+//   - Spartan API: expecting users to plainly use sync.Cond is risky, as the API
+//     is fairly low-level.
+//   - No solution for late consumers: late consumers (ones that missed the value
+//     being set by a producer) would still have to implement logic in order to
+//     find out such a value, as sync.Cond only supports what amounts to
+//     edge-level triggers as part of its Broadcast/Signal system.
+//
+// It would be possible to implement MemoryValue using a sync.Cond internally,
+// but such an implementation would likely be more complex than the current
+// implementation based on channels and mutexes, as it would have to work
+// around issues like lack of canceling, etc.
+package event
+
+import (
+	"context"
+	"errors"
+
+	"source.monogon.dev/osbase/supervisor"
+)
+
+// A Value is an 'Event Value', some piece of data that can be updated ('Set')
+// by Producers and retrieved by Consumers.
+type Value[T any] interface {
+	// Set updates the Value to the given data. It is safe to call this from
+	// multiple goroutines, including concurrently.
+	//
+	// Any time Set is called, any consumers performing a Watch on this Value
+	// will be notified with the new data - even if the Set data is the same as
+	// the one that was already stored.
+	//
+	// A Value will initially have no data set. This 'no data' state is seen by
+	// consumers by the first .Get() call on the Watcher blocking until data is Set.
+	//
+	// All updates will be serialized in an arbitrary order - if multiple
+	// producers wish to perform concurrent actions to update the Value partially,
+	// this should be negotiated and serialized externally by the producers.
+	Set(val T)
+
+	// ValueWatch implements the Watch method. It is split out into another
+	// interface to allow some 'Event Values' to implement only the watch/read
+	// part, with the write side being implicit or defined by a more complex
+	// interface than a simple Set().
+	ValueWatch[T]
+}
+
+// ValueWatch is the read side of an 'Event Value', witch can by retrieved by
+// Consumers by performing a Watch operation on it.
+type ValueWatch[T any] interface {
+	// Watch retrieves a Watcher that keeps track on the version of the data
+	// contained within the Value that was last seen by a consumer. Once a
+	// Watcher is retrieved, it can be used to then get the actual data stored
+	// within the Value, and to reliably retrieve updates to it without having
+	// to poll for changes.
+	Watch() Watcher[T]
+}
+
+// A Watcher keeps track of the last version of data seen by a consumer for a
+// given Value. Each consumer should use an own Watcher instance, and it is not
+// safe to use this type concurrently. However, it is safe to move/copy it
+// across different goroutines, as long as no two goroutines access it
+// simultaneously.
+type Watcher[T any] interface {
+	// Get blocks until a Value's data is available:
+	//  - On first use of a Watcher, Get will return the data contained in the
+	//    value at the time of calling .Watch(), or block if no data has been
+	//    .Set() on it yet. If a value has been Set() since the initial
+	//    creation of the Watch() but before Get() is called for the first
+	//    time, the first Get() call will immediately return the new value.
+	//  - On subsequent uses of a Watcher, Get will block until the given Value
+	//    has been Set with new data. This does not necessarily mean that the
+	//    new data is different - consumers should always perform their own
+	//    checks on whether the update is relevant to them (ie., the data has
+	//    changed in a significant way), unless specified otherwise by a Value
+	//    publisher.
+	//
+	// Get() will always return the current newest data that has been Set() on
+	// the Value, and not a full log of historical events. This is geared
+	// towards event values where consumers only care about changes to data
+	// since last retrieval, not every value that has been Set along the way.
+	// Thus, consumers need not make sure that they actively .Get() on a
+	// watcher all the times.
+	//
+	// If the context is canceled before data is available to be returned, the
+	// context's error will be returned. However, the Watcher will still need to be
+	// Closed, as it is still fully functional after the context has been canceled.
+	//
+	// Concurrent requests to Get result in an error. The reasoning to return
+	// an error instead of attempting to serialize the requests is that any
+	// concurrent access from multiple goroutines would cause a desync in the
+	// next usage of the Watcher. For example:
+	//   1) w.Get() (in G0) and w.Get(G1) start. They both block waiting for an
+	//      initial value.
+	//   2) v.Set(0)
+	//   3) w.Get() in G0 returns 0,
+	//   4) v.Set(1)
+	//   4) w.Get() in G1 returns 1,
+	// This would cause G0 and G1 to become desynchronized between eachother
+	// (both have different value data) and subsequent updates will also
+	// continue skipping some updates.
+	// If multiple goroutines need to access the Value, they should each use
+	// their own Watcher.
+	Get(context.Context, ...GetOption[T]) (T, error)
+
+	// Close must be called if the Watcher is not going to be used anymore -
+	// otherwise, it will not be garbage collected.
+	Close() error
+}
+
+type GetOption[T any] struct {
+	Predicate   func(t T) bool
+	BacklogOnly bool
+}
+
+func Filter[T any](pred func(T) bool) GetOption[T] {
+	return GetOption[T]{
+		Predicate: pred,
+	}
+}
+
+// BacklogOnly will prevent Get from blocking on waiting for more updates from
+// etcd, by instead returning ErrBacklogDone whenever no more data is currently
+// locally available. This is different however, from establishing that there
+// are no more pending updates from the etcd cluster - the only way to ensure
+// the local client is up to date is by performing Get calls without this option
+// set.
+//
+// This mode of retrieval should only be used for the retrieval of the existing
+// data in the etcd cluster on the initial creation of the Watcher (by
+// repeatedly calling Get until ErrBacklogDone is returned), and shouldn't be set
+// for any subsequent call. Any use of this option after that initial fetch is
+// undefined behaviour that exposes the internals of the Get implementation, and
+// must not be relied on. However, in the future, this behaviour might be
+// formalized.
+//
+// This mode is particularly useful for ranged watchers. Non-ranged watchers can
+// still use this option to distinguish between blocking because of the
+// nonexistence of an object vs. blocking because of networking issues. However,
+// non-ranged retrieval semantics generally will rarely need to make this
+// distinction.
+func BacklogOnly[T any]() GetOption[T] {
+	return GetOption[T]{BacklogOnly: true}
+}
+
+var (
+	// ErrBacklogDone is returned by Get when BacklogOnly is set and there is no more
+	// event data stored in the Watcher client, ie. when the initial cluster state
+	// of the requested key has been retrieved.
+	ErrBacklogDone = errors.New("no more backlogged data")
+)
+
+// Pipe a Value's initial state and subsequent updates to an already existing
+// channel in a supervisor.Runnable. This is mostly useful when wanting to select
+// {} on many Values.
+//
+// The given channel will NOT be closed when the runnable exits. The process
+// receiving from the channel should be running in a group with the pipe
+// runnable, so that both restart if either does. This ensures that there is always
+// at least one value in the channel when the receiver starts.
+func Pipe[T any](value Value[T], c chan<- T, opts ...GetOption[T]) supervisor.Runnable {
+	return func(ctx context.Context) error {
+		supervisor.Signal(ctx, supervisor.SignalHealthy)
+		w := value.Watch()
+		defer w.Close()
+		for {
+			v, err := w.Get(ctx, opts...)
+			if err != nil {
+				return err
+			}
+			select {
+			case c <- v:
+			case <-ctx.Done():
+				return ctx.Err()
+			}
+		}
+	}
+}
diff --git a/osbase/event/memory/BUILD.bazel b/osbase/event/memory/BUILD.bazel
new file mode 100644
index 0000000..f2cd4bd
--- /dev/null
+++ b/osbase/event/memory/BUILD.bazel
@@ -0,0 +1,19 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "memory",
+    srcs = ["memory.go"],
+    importpath = "source.monogon.dev/osbase/event/memory",
+    visibility = ["//visibility:public"],
+    deps = ["//osbase/event"],
+)
+
+go_test(
+    name = "memory_test",
+    srcs = [
+        "example_test.go",
+        "memory_test.go",
+    ],
+    embed = [":memory"],
+    deps = ["//osbase/event"],
+)
diff --git a/osbase/event/memory/example_test.go b/osbase/event/memory/example_test.go
new file mode 100644
index 0000000..1ae12c6
--- /dev/null
+++ b/osbase/event/memory/example_test.go
@@ -0,0 +1,114 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memory
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"time"
+)
+
+// NetworkStatus is example data that will be stored in a Value.
+type NetworkStatus struct {
+	ExternalAddress net.IP
+	DefaultGateway  net.IP
+}
+
+// NetworkService is a fake/example network service that is responsible for
+// communicating the newest information about a machine's network configuration
+// to consumers/watchers.
+type NetworkService struct {
+	Provider Value[NetworkStatus]
+}
+
+// Run pretends to execute the network service's main logic loop, in which it
+// pretends to have received an IP address over DHCP, and communicates that to
+// consumers/watchers.
+func (s *NetworkService) Run(ctx context.Context) {
+	s.Provider.Set(NetworkStatus{
+		ExternalAddress: nil,
+		DefaultGateway:  nil,
+	})
+
+	select {
+	case <-time.After(100 * time.Millisecond):
+	case <-ctx.Done():
+		return
+	}
+
+	fmt.Printf("NS: Got DHCP Lease\n")
+	s.Provider.Set(NetworkStatus{
+		ExternalAddress: net.ParseIP("203.0.113.24"),
+		DefaultGateway:  net.ParseIP("203.0.113.1"),
+	})
+
+	select {
+	case <-time.After(100 * time.Millisecond):
+	case <-ctx.Done():
+		return
+	}
+
+	fmt.Printf("NS: DHCP Address changed\n")
+	s.Provider.Set(NetworkStatus{
+		ExternalAddress: net.ParseIP("203.0.113.103"),
+		DefaultGateway:  net.ParseIP("203.0.113.1"),
+	})
+
+	time.Sleep(100 * time.Millisecond)
+}
+
+// ExampleValue_full demonstrates a typical usecase for Event Values, in which
+// a mock network service lets watchers know that the machine on which the code
+// is running has received a new network configuration.
+// It also shows the typical boilerplate required in order to wrap a Value (eg.
+// MemoryValue) within a typesafe wrapper.
+func ExampleValue_full() {
+	ctx, ctxC := context.WithCancel(context.Background())
+	defer ctxC()
+
+	// Create a fake NetworkService.
+	var ns NetworkService
+
+	// Run an /etc/hosts updater. It will watch for updates from the NetworkService
+	// about the current IP address of the node.
+	go func() {
+		w := ns.Provider.Watch()
+		for {
+			status, err := w.Get(ctx)
+			if err != nil {
+				break
+			}
+			if status.ExternalAddress == nil {
+				continue
+			}
+			// Pretend to write /etc/hosts with the newest ExternalAddress.
+			// In production code, you would also check for whether ExternalAddress has
+			// changed from the last written value, if writing to /etc/hosts is expensive.
+			fmt.Printf("/etc/hosts: foo.example.com is now %s\n", status.ExternalAddress.String())
+		}
+	}()
+
+	// Run fake network service.
+	ns.Run(ctx)
+
+	// Output:
+	// NS: Got DHCP Lease
+	// /etc/hosts: foo.example.com is now 203.0.113.24
+	// NS: DHCP Address changed
+	// /etc/hosts: foo.example.com is now 203.0.113.103
+}
diff --git a/osbase/event/memory/memory.go b/osbase/event/memory/memory.go
new file mode 100644
index 0000000..16818a0
--- /dev/null
+++ b/osbase/event/memory/memory.go
@@ -0,0 +1,233 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memory
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"sync"
+
+	"source.monogon.dev/osbase/event"
+)
+
+var (
+	// Type assert that *Value implements Value. We do this artificially, as
+	// there currently is no code path that needs this to be strictly true. However,
+	// users of this library might want to rely on the Value type instead of
+	// particular Value implementations.
+	_ event.Value[int] = &Value[int]{}
+)
+
+// Value is a 'memory value', which implements a event.Value stored in memory.
+// It is safe to construct an empty object of this type. However, this must not
+// be copied.
+type Value[T any] struct {
+	// mu guards the inner, innerSet and watchers fields.
+	mu sync.RWMutex
+	// inner is the latest data Set on the Value. It is used to provide the
+	// newest version of the Set data to new watchers.
+	inner T
+	// innerSet is true when inner has been Set at least once. It is used to
+	// differentiate between a nil and unset value.
+	innerSet bool
+	// watchers is the list of watchers that should be updated when new data is
+	// Set. It will grow on every .Watch() and shrink any time a watcher is
+	// determined to have been closed.
+	watchers []*watcher[T]
+
+	// Sync, if set to true, blocks all .Set() calls on the Value until all
+	// Watchers derived from it actively .Get() the new value. This can be used
+	// to ensure Watchers always receive a full log of all Set() calls.
+	//
+	// This must not be changed after the first .Set/.Watch call.
+	//
+	// This is an experimental API and subject to change. It might be migrated
+	// to per-Watcher settings defined within the main event.Value/Watcher
+	// interfaces.
+	Sync bool
+}
+
+// Set updates the Value to the given data. It is safe to call this from
+// multiple goroutines, including concurrently.
+//
+// For more information about guarantees, see event.Value.Set.
+func (m *Value[T]) Set(val T) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	// Update the data that is provided on first Get() to watchers.
+	m.inner = val
+	m.innerSet = true
+
+	// Go through all watchers, updating them on the new value and filtering out
+	// all closed watchers.
+	newWatchers := m.watchers[:0]
+	for _, w := range m.watchers {
+		if w.closed() {
+			continue
+		}
+		w.update(m.Sync, val)
+		newWatchers = append(newWatchers, w)
+	}
+	if cap(newWatchers) > len(newWatchers)*3 {
+		reallocated := make([]*watcher[T], 0, len(newWatchers)*2)
+		newWatchers = append(reallocated, newWatchers...)
+	}
+	m.watchers = newWatchers
+}
+
+// watcher implements the event.Watcher interface for watchers returned by
+// Value.
+type watcher[T any] struct {
+	// bufferedC is a buffered channel of size 1 for submitting values to the
+	// watcher.
+	bufferedC chan T
+	// unbufferedC is an unbuffered channel, which is used when Sync is enabled.
+	unbufferedC chan T
+
+	// getSem is a channel-based semaphore (which is of size 1, and thus in
+	// fact a mutex) that is used to ensure that only a single .Get() call is
+	// active. It is implemented as a channel to permit concurrent .Get() calls
+	// to error out instead of blocking.
+	getSem chan struct{}
+	// close is a channel that is closed when this watcher is itself Closed.
+	close chan struct{}
+}
+
+// Watch retrieves a Watcher that keeps track on the version of the data
+// contained within the Value that was last seen by a consumer.
+//
+// For more information about guarantees, see event.Value.Watch.
+func (m *Value[T]) Watch() event.Watcher[T] {
+	waiter := &watcher[T]{
+		bufferedC:   make(chan T, 1),
+		unbufferedC: make(chan T),
+		close:       make(chan struct{}),
+		getSem:      make(chan struct{}, 1),
+	}
+
+	m.mu.Lock()
+	// If the watchers slice is at capacity, drop closed watchers, and
+	// reallocate the slice at 2x length if it is not between 1.5x and 3x.
+	if len(m.watchers) == cap(m.watchers) {
+		newWatchers := m.watchers[:0]
+		for _, w := range m.watchers {
+			if !w.closed() {
+				newWatchers = append(newWatchers, w)
+			}
+		}
+		if cap(newWatchers)*2 < len(newWatchers)*3 || cap(newWatchers) > len(newWatchers)*3 {
+			reallocated := make([]*watcher[T], 0, len(newWatchers)*2)
+			newWatchers = append(reallocated, newWatchers...)
+		}
+		m.watchers = newWatchers
+	}
+	// Append this watcher to the Value.
+	m.watchers = append(m.watchers, waiter)
+	// If the Value already has some value set, put it in the buffered channel.
+	if m.innerSet {
+		waiter.bufferedC <- m.inner
+	}
+	m.mu.Unlock()
+
+	return waiter
+}
+
+// closed returns whether this watcher has been closed.
+func (m *watcher[T]) closed() bool {
+	select {
+	case _, ok := <-m.close:
+		if !ok {
+			return true
+		}
+	default:
+	}
+	return false
+}
+
+// update is the high level update-this-watcher function called by Value.
+func (m *watcher[T]) update(sync bool, val T) {
+	// If synchronous delivery was requested, block until a watcher .Gets it,
+	// or is closed.
+	if sync {
+		select {
+		case m.unbufferedC <- val:
+		case <-m.close:
+		}
+		return
+	}
+
+	// Otherwise, deliver asynchronously. If there is already a value in the
+	// buffered channel that was not retrieved, drop it.
+	select {
+	case <-m.bufferedC:
+	default:
+	}
+	// The channel is now empty, so sending to it cannot block.
+	m.bufferedC <- val
+}
+
+func (m *watcher[T]) Close() error {
+	close(m.close)
+	return nil
+}
+
+// Get blocks until a Value's data is available. See event.Watcher.Get for
+// guarantees and more information.
+func (m *watcher[T]) Get(ctx context.Context, opts ...event.GetOption[T]) (T, error) {
+	// Make sure we're the only active .Get call.
+	var empty T
+	select {
+	case m.getSem <- struct{}{}:
+	default:
+		return empty, fmt.Errorf("cannot Get() concurrently on a single waiter")
+	}
+	defer func() {
+		<-m.getSem
+	}()
+
+	var predicate func(t T) bool
+	for _, opt := range opts {
+		if opt.Predicate != nil {
+			predicate = opt.Predicate
+		}
+		if opt.BacklogOnly {
+			return empty, errors.New("BacklogOnly is not implemented for memory watchers")
+		}
+	}
+
+	for {
+		var val T
+		// For Sync values, ensure the initial value in the buffered
+		// channel is delivered first.
+		select {
+		case val = <-m.bufferedC:
+		default:
+			select {
+			case <-ctx.Done():
+				return empty, ctx.Err()
+			case val = <-m.bufferedC:
+			case val = <-m.unbufferedC:
+			}
+		}
+		if predicate != nil && !predicate(val) {
+			continue
+		}
+		return val, nil
+	}
+}
diff --git a/osbase/event/memory/memory_test.go b/osbase/event/memory/memory_test.go
new file mode 100644
index 0000000..b622565
--- /dev/null
+++ b/osbase/event/memory/memory_test.go
@@ -0,0 +1,371 @@
+// Copyright 2020 The Monogon Project Authors.
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memory
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"source.monogon.dev/osbase/event"
+)
+
+// TestAsync exercises the high-level behaviour of a Value, in which a
+// watcher is able to catch up to the newest Set value.
+func TestAsync(t *testing.T) {
+	p := Value[int]{}
+	p.Set(0)
+
+	ctx := context.Background()
+
+	// The 0 from Set() should be available via .Get().
+	watcher := p.Watch()
+	val, err := watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+	if want, got := 0, val; want != got {
+		t.Fatalf("Value: got %d, wanted %d", got, want)
+	}
+
+	// Send a large amount of updates that the watcher does not actively .Get().
+	for i := 1; i <= 100; i++ {
+		p.Set(i)
+	}
+
+	// The watcher should still end up with the newest .Set() value on the next
+	// .Get() call.
+	val, err = watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+	if want, got := 100, val; want != got {
+		t.Fatalf("Value: got %d, wanted %d", got, want)
+	}
+}
+
+// TestSyncBlocks exercises the Value's 'Sync' field, which makes all
+// Set() calls block until all respective watchers .Get() the updated data.
+// This particular test ensures that .Set() calls to a Watcher result in a
+// prefect log of updates being transmitted to a watcher.
+func TestSync(t *testing.T) {
+	p := Value[int]{
+		Sync: true,
+	}
+	values := make(chan int, 100)
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() {
+		ctx := context.Background()
+		watcher := p.Watch()
+		wg.Done()
+		for {
+			value, err := watcher.Get(ctx)
+			if err != nil {
+				panic(err)
+			}
+			values <- value
+		}
+	}()
+
+	p.Set(0)
+	wg.Wait()
+
+	want := []int{1, 2, 3, 4}
+	for _, w := range want {
+		p.Set(w)
+	}
+
+	timeout := time.After(time.Second)
+	for i, w := range append([]int{0}, want...) {
+		select {
+		case <-timeout:
+			t.Fatalf("timed out on value %d (%d)", i, w)
+		case val := <-values:
+			if w != val {
+				t.Errorf("value %d was %d, wanted %d", i, val, w)
+			}
+		}
+	}
+}
+
+// TestSyncBlocks exercises the Value's 'Sync' field, which makes all
+// Set() calls block until all respective watchers .Get() the updated data.
+// This particular test ensures that .Set() calls actually block when a watcher
+// is unattended.
+func TestSyncBlocks(t *testing.T) {
+	p := Value[int]{
+		Sync: true,
+	}
+	ctx := context.Background()
+
+	// Shouldn't block, as there's no declared watchers.
+	p.Set(0)
+
+	watcher := p.Watch()
+
+	// Should retrieve the zero, more requests will pend.
+	value, err := watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+	if want, got := 0, value; want != got {
+		t.Fatalf("Got initial value %d, wanted %d", got, want)
+	}
+
+	// .Set() Should block, as watcher is unattended.
+	//
+	// Whether something blocks in Go is untestable in a robust way (see: halting
+	// problem). We work around this this by introducing a 'stage' int64, which is
+	// put on the 'c' channel after the needs-to-block function returns. We then
+	// perform an action that should unblock this function right after updating
+	// 'stage' to a different value.
+	// Then, we observe what was put on the channel: If it's the initial value, it
+	// means the function didn't block when expected. Otherwise, it means the
+	// function unblocked when expected.
+	stage := int64(0)
+	c := make(chan int64, 1)
+	go func() {
+		p.Set(1)
+		c <- atomic.LoadInt64(&stage)
+	}()
+
+	// Getting should unblock the provider. Mark via 'stage' variable that
+	// unblocking now is expected.
+	atomic.StoreInt64(&stage, int64(1))
+	// Potential race: .Set() unblocks here due to some bug, before .Get() is
+	// called, and we record a false positive.
+	value, err = watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+
+	res := <-c
+	if res != int64(1) {
+		t.Fatalf("Set() returned before Get()")
+	}
+
+	if want, got := 1, value; want != got {
+		t.Fatalf("Wanted value %d, got %d", want, got)
+	}
+
+	// Closing the watcher and setting should not block anymore.
+	if err := watcher.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	// Last step, if this blocks we will get a deadlock error and the test will panic.
+	p.Set(2)
+}
+
+// TestMultipleGets verifies that calling .Get() on a single watcher from two
+// goroutines is prevented by returning an error in exactly one of them.
+func TestMultipleGets(t *testing.T) {
+	p := Value[int]{}
+	ctx := context.Background()
+
+	w := p.Watch()
+
+	tryError := func(errs chan error) {
+		_, err := w.Get(ctx)
+		errs <- err
+	}
+	errs := make(chan error, 2)
+	go tryError(errs)
+	go tryError(errs)
+
+	for err := range errs {
+		if err == nil {
+			t.Fatalf("A Get call succeeded, while it should have blocked or returned an error")
+		} else {
+			// Found the error, test succeeded.
+			break
+		}
+	}
+}
+
+// TestConcurrency attempts to stress the Value/Watcher
+// implementation to design limits (a hundred simultaneous watchers), ensuring
+// that the watchers all settle to the final set value.
+func TestConcurrency(t *testing.T) {
+	ctx := context.Background()
+
+	p := Value[int]{}
+	p.Set(0)
+
+	// Number of watchers to create.
+	watcherN := 100
+	// Expected final value to be Set().
+	final := 100
+	// Result channel per watcher.
+	resC := make([]chan error, watcherN)
+
+	// Spawn watcherN watchers.
+	for i := 0; i < watcherN; i++ {
+		resC[i] = make(chan error, 1)
+		go func(id int) {
+			// done is a helper function that will put an error on the
+			// respective watcher's resC.
+			done := func(err error) {
+				resC[id] <- err
+				close(resC[id])
+			}
+
+			watcher := p.Watch()
+			// prev is used to ensure the values received are monotonic.
+			prev := -1
+			for {
+				val, err := watcher.Get(ctx)
+				if err != nil {
+					done(err)
+					return
+				}
+
+				// Ensure monotonicity of received data.
+				if val <= prev {
+					done(fmt.Errorf("received out of order data: %d after %d", val, prev))
+				}
+				prev = val
+
+				// Quit when the final value is received.
+				if val == final {
+					done(nil)
+					return
+				}
+
+				// Sleep a bit, depending on the watcher. This makes each
+				// watcher behave slightly differently, and attempts to
+				// exercise races dependent on sleep time between subsequent
+				// Get calls.
+				time.Sleep(time.Millisecond * time.Duration(id))
+			}
+		}(i)
+	}
+
+	// Set 1..final on the value.
+	for i := 1; i <= final; i++ {
+		p.Set(i)
+	}
+
+	// Ensure all watchers exit with no error.
+	for i, c := range resC {
+		err := <-c
+		if err != nil {
+			t.Errorf("Watcher %d returned %v", i, err)
+		}
+	}
+}
+
+// TestCanceling exercises whether a context canceling in a .Get() gracefully
+// aborts that particular Get call, but also allows subsequent use of the same
+// watcher.
+func TestCanceling(t *testing.T) {
+	p := Value[int]{
+		Sync: true,
+	}
+
+	ctx, ctxC := context.WithCancel(context.Background())
+
+	watcher := p.Watch()
+
+	// errs will contain the error returned by Get.
+	errs := make(chan error, 1)
+	go func() {
+		// This Get will block, as no initial data has been Set on the value.
+		_, err := watcher.Get(ctx)
+		errs <- err
+	}()
+
+	// Cancel the context, and expect that context error to propagate to the .Get().
+	ctxC()
+	if want, got := ctx.Err(), <-errs; !errors.Is(got, want) {
+		t.Fatalf("Get should've returned %v, got %v", want, got)
+	}
+
+	// Do another .Get() on the same watcher with a new context. Even though the
+	// call was aborted via a context cancel, the watcher should continue working.
+	ctx = context.Background()
+	go func() {
+		_, err := watcher.Get(ctx)
+		errs <- err
+	}()
+
+	// Unblock the .Get now.
+	p.Set(1)
+	if want, got := error(nil), <-errs; !errors.Is(got, want) {
+		t.Fatalf("Get should've returned %v, got %v", want, got)
+	}
+}
+
+// TestSetAfterWatch ensures that if a value is updated between a Watch and the
+// initial Get, only the newest Set value is returns.
+func TestSetAfterWatch(t *testing.T) {
+	ctx := context.Background()
+
+	p := Value[int]{}
+	p.Set(0)
+
+	watcher := p.Watch()
+	p.Set(1)
+
+	data, err := watcher.Get(ctx)
+	if err != nil {
+		t.Fatalf("Get: %v", err)
+	}
+	if want, got := 1, data; want != got {
+		t.Errorf("Get should've returned %v, got %v", want, got)
+	}
+}
+
+// TestWatchersList ensures that the list of watchers is managed correctly,
+// i.e. there is no memory leak and closed watchers are removed while
+// keeping all non-closed watchers.
+func TestWatchersList(t *testing.T) {
+	ctx := context.Background()
+	p := Value[int]{}
+
+	var watchers []event.Watcher[int]
+	for i := 0; i < 100; i++ {
+		watchers = append(watchers, p.Watch())
+	}
+	for i := 0; i < 10000; i++ {
+		watchers[10].Close()
+		watchers[10] = p.Watch()
+	}
+
+	if want, got := 1000, cap(p.watchers); want <= got {
+		t.Fatalf("Got capacity %d, wanted less than %d", got, want)
+	}
+
+	p.Set(1)
+	if want, got := 100, len(p.watchers); want != got {
+		t.Fatalf("Got %d watchers, wanted %d", got, want)
+	}
+
+	for _, watcher := range watchers {
+		data, err := watcher.Get(ctx)
+		if err != nil {
+			t.Fatalf("Get: %v", err)
+		}
+		if want, got := 1, data; want != got {
+			t.Errorf("Get should've returned %v, got %v", want, got)
+		}
+	}
+}