diff --git a/include/dqlite.h b/include/dqlite.h
index f064f4836..d0dbe153f 100644
--- a/include/dqlite.h
+++ b/include/dqlite.h
@@ -586,6 +586,29 @@ DQLITE_API int dqlite_node_recover_ext(dqlite_node *n,
 				       dqlite_node_info_ext infos[],
 				       int n_info);
 
+/**
+ * Retrieve information about the last persisted raft log entry.
+ *
+ * This is intended to be used in combination with dqlite_node_recover_ext, to
+ * determine which of the surviving nodes in a cluster is most up-to-date. The
+ * raft rules for this are:
+ *
+ * - If the two logs have last entries with different terms, the log with the
+ *   higher term is more up-to-date.
+ * - Otherwise, the longer log is more up-to-date.
+ *
+ * Note that this function may result in physically modifying the raft-related
+ * files in the data directory. These modifications do not affect the logical
+ * state of the node. Deletion of invalid segment files can be disabled with
+ * dqlite_node_set_auto_recovery.
+ *
+ * This should be called after dqlite_node_init, but the node must not be
+ * running.
+ */
+DQLITE_API int dqlite_node_describe_last_entry(dqlite_node *n,
+					       uint64_t *last_entry_index,
+					       uint64_t *last_entry_term);
+
 /**
  * Return a human-readable description of the last error occurred.
  */
diff --git a/src/raft.h b/src/raft.h
index fce3f5518..76bea3500 100644
--- a/src/raft.h
+++ b/src/raft.h
@@ -1061,6 +1061,28 @@ RAFT_API int raft_bootstrap(struct raft *r,
 RAFT_API int raft_recover(struct raft *r,
 			  const struct raft_configuration *conf);
 
+/**
+ * Read information about the last raft log entry that's stored on disk.
+ *
+ * "Last log entry" here should be understood as including snapshots,
+ * so if there is one snapshot on disk and no individual entries, the
+ * values returned in `index` and `term` are the index and term of the
+ * last entry included in the snapshot. If there are no snapshot and no
+ * entries, then `index` and `term` are both set to 0.
+ *
+ * This function is just a wrapper around the `load` method of raft_io.
+ * Note that the `load` method of the uv raft_io implementation is not
+ * read-only: as it walks the segment files on disk, it closes open
+ * segments that contain valid entries and deletes other open segments.
+ *
+ * This should be called after the raft_io instance is initialized (e.g.
+ * after calling raft_uv_init), but no active raft node should be using
+ * the instance.
+ */
+int raft_io_describe_last_entry(struct raft_io *io,
+				raft_index *index,
+				raft_term *term);
+
 RAFT_API int raft_start(struct raft *r);
 
 /**
diff --git a/src/raft/raft.c b/src/raft/raft.c
index ce3996b90..e98927b4e 100644
--- a/src/raft/raft.c
+++ b/src/raft/raft.c
@@ -10,11 +10,13 @@
 #include "configuration.h"
 #include "convert.h"
 #include "election.h"
+#include "entry.h"
 #include "err.h"
 #include "flags.h"
 #include "heap.h"
 #include "log.h"
 #include "membership.h"
+#include "snapshot.h"
 
 #define DEFAULT_ELECTION_TIMEOUT 1000          /* One second */
 #define DEFAULT_HEARTBEAT_TIMEOUT 100          /* One tenth of a second */
@@ -303,3 +305,31 @@ static int ioFsmVersionCheck(struct raft *r,
 
 	return 0;
 }
+
+int raft_io_describe_last_entry(struct raft_io *io,
+				raft_index *index,
+				raft_term *term)
+{
+	raft_term current_term;
+	raft_id voted_for;
+	struct raft_snapshot *snapshot;
+	raft_index start_index;
+	struct raft_entry *entries;
+	size_t n_entries;
+	int rv;
+
+	rv = io->load(io, &current_term, &voted_for, &snapshot,
+		      &start_index, &entries, &n_entries);
+	if (rv != 0) {
+		return rv;
+	}
+	*index = start_index + n_entries - 1;
+	*term = n_entries > 0 ? entries[n_entries - 1].term :
+		snapshot != NULL ? snapshot->term :
+		0;
+	if (snapshot != NULL) {
+		snapshotDestroy(snapshot);
+	}
+	entryBatchesDestroy(entries, n_entries);
+	return 0;
+}
diff --git a/src/server.c b/src/server.c
index 7e44b1c49..6a29e80c3 100644
--- a/src/server.c
+++ b/src/server.c
@@ -1065,6 +1065,30 @@ int dqlite_node_recover_ext(dqlite_node *n,
 	return rv;
 }
 
+int dqlite_node_describe_last_entry(dqlite_node *n,
+				    uint64_t *index,
+				    uint64_t *term)
+{
+	PRE(n->initialized && !n->running);
+	static_assert(sizeof(*index) == sizeof(raft_index),
+		      "unexpected index type size");
+	raft_index *i = (raft_index *)index;
+	static_assert(sizeof(*term) == sizeof(raft_term),
+		      "unexpected term type size");
+	raft_term *t = (raft_term *)term;
+	int rv;
+
+#ifdef USE_SYSTEM_RAFT
+	(void)i;
+	(void)t;
+	(void)rv;
+	return DQLITE_ERROR;
+#else
+	rv = raft_io_describe_last_entry(&n->raft_io, i, t);
+	return rv == 0 ? 0 : DQLITE_ERROR;
+#endif
+}
+
 dqlite_node_id dqlite_generate_node_id(const char *address)
 {
 	tracef("generate node id");
diff --git a/test/integration/test_cluster.c b/test/integration/test_cluster.c
index 57f605cff..131be5132 100644
--- a/test/integration/test_cluster.c
+++ b/test/integration/test_cluster.c
@@ -97,17 +97,16 @@ TEST(cluster, restart, setUp, tearDown, 0, cluster_params)
 	struct rows rows;
 	long n_records =
 	    strtol(munit_parameters_get(params, "num_records"), NULL, 0);
-	char sql[128];
 
 	HANDSHAKE;
 	OPEN;
 	PREPARE("CREATE TABLE test (n INT)", &stmt_id);
 	EXEC(stmt_id, &last_insert_id, &rows_affected);
 
+	PREPARE("INSERT INTO TEST(n) VALUES(?)", &stmt_id);
 	for (int i = 0; i < n_records; ++i) {
-		sprintf(sql, "INSERT INTO test(n) VALUES(%d)", i + 1);
-		PREPARE(sql, &stmt_id);
-		EXEC(stmt_id, &last_insert_id, &rows_affected);
+		EXEC_PARAMS(stmt_id, &last_insert_id, &rows_affected,
+			    {.type = SQLITE_INTEGER, .integer = i});
 	}
 
 	struct test_server *server = &f->servers[0];
@@ -133,7 +132,6 @@ TEST(cluster, dataOnNewNode, setUp, tearDown, 0, cluster_params)
 	struct rows rows;
 	long n_records =
 	    strtol(munit_parameters_get(params, "num_records"), NULL, 0);
-	char sql[128];
 	unsigned id = 2;
 	const char *address = "@2";
 
@@ -142,10 +140,10 @@ TEST(cluster, dataOnNewNode, setUp, tearDown, 0, cluster_params)
 	PREPARE("CREATE TABLE test (n INT)", &stmt_id);
 	EXEC(stmt_id, &last_insert_id, &rows_affected);
 
+	PREPARE("INSERT INTO test(n) VALUES(?)", &stmt_id);
 	for (int i = 0; i < n_records; ++i) {
-		sprintf(sql, "INSERT INTO test(n) VALUES(%d)", i + 1);
-		PREPARE(sql, &stmt_id);
-		EXEC(stmt_id, &last_insert_id, &rows_affected);
+		EXEC_PARAMS(stmt_id, &last_insert_id, &rows_affected,
+			    {.type = SQLITE_INTEGER, .integer = i});
 	}
 
 	/* Add a second voting server, this one will receive all data from the
@@ -158,6 +156,32 @@ TEST(cluster, dataOnNewNode, setUp, tearDown, 0, cluster_params)
 	REMOVE(1);
 	sleep(1);
 
+	struct test_server *first = &f->servers[0];
+	test_server_stop(first);
+	test_server_prepare(first, params);
+#ifndef USE_SYSTEM_RAFT
+	int rv;
+	/* One entry per INSERT, plus one for the initial configuration, plus
+	 * one for the CREATE TABLE, plus one legacy checkpoint command entry
+	 * after 993 records or two after 2200 records. */
+	uint64_t expected_entries = n_records + (n_records >= 2200 ? 4 :
+						 n_records >= 993 ? 3 :
+						 2);
+	/* We also expect a variable number of barrier entries. Just specify an
+	 * upper bound since we don't know the exact count. */
+	uint64_t max_barriers = 10;
+	uint64_t last_entry_index;
+	uint64_t last_entry_term;
+	rv = dqlite_node_describe_last_entry(first->dqlite,
+					     &last_entry_index,
+					     &last_entry_term);
+	munit_assert_int(rv, ==, 0);
+	munit_assert_uint64(expected_entries, <=, last_entry_index);
+	munit_assert_uint64(last_entry_index, <, expected_entries + max_barriers);
+	munit_assert_uint64(last_entry_term, ==, 1);
+#endif
+	test_server_run(first);
+
 	/* The full table is visible from the new node */
 	SELECT(2);
 	HANDSHAKE;
@@ -166,6 +190,25 @@ TEST(cluster, dataOnNewNode, setUp, tearDown, 0, cluster_params)
 	QUERY(stmt_id, &rows);
 	munit_assert_long(rows.next->values->integer, ==, n_records);
 	clientCloseRows(&rows);
+
+	/* One more entry on the new node. */
+	PREPARE("INSERT INTO test(n) VALUES(?)", &stmt_id);
+	EXEC_PARAMS(stmt_id, &last_insert_id, &rows_affected,
+		    {.type = SQLITE_INTEGER, .integer = 5000});
+
+	struct test_server *second = &f->servers[1];
+	test_server_stop(second);
+	test_server_prepare(second, params);
+#ifndef USE_SYSTEM_RAFT
+	rv = dqlite_node_describe_last_entry(second->dqlite,
+					     &last_entry_index,
+					     &last_entry_term);
+	munit_assert_int(rv, ==, 0);
+	munit_assert_uint64(expected_entries + 1, <=, last_entry_index);
+	munit_assert_uint64(last_entry_index, <, expected_entries + max_barriers + 1);
+	munit_assert_uint64(last_entry_term, ==, 1);
+#endif
+	test_server_run(second);
 	return MUNIT_OK;
 }
 
@@ -288,3 +331,42 @@ TEST(cluster, modifyingQuerySql, setUp, tearDown, 0, cluster_params)
 	clientCloseRows(&rows);
 	return MUNIT_OK;
 }
+
+#ifndef USE_SYSTEM_RAFT
+
+/* Edge cases for dqlite_node_describe_last_entry. */
+TEST(cluster, last_entry_edge_cases, setUp, tearDown, 0, NULL)
+{
+	struct fixture *f = data;
+	uint64_t index;
+	uint64_t term;
+	int rv;
+
+	sleep(1);
+
+	struct test_server *first = &f->servers[0];
+	test_server_stop(first);
+	test_server_prepare(first, params);
+	rv = dqlite_node_describe_last_entry(first->dqlite, &index, &term);
+	munit_assert_int(rv, ==, 0);
+	/* The log contains only the bootstrap configuration. */
+	munit_assert_uint64(index, ==, 1);
+	/* The bootstrap configuration is always tagged with term 1. */
+	munit_assert_uint64(term, ==, 1);
+	test_server_run(first);
+
+	struct test_server *second = &f->servers[1];
+	test_server_stop(second);
+	test_server_prepare(second, params);
+	rv = dqlite_node_describe_last_entry(second->dqlite, &index, &term);
+	munit_assert_int(rv, ==, 0);
+	/* We didn't bootstrap and haven't joined the leader, so our log is
+	 * empty. */
+	munit_assert_uint64(index, ==, 0);
+	munit_assert_uint64(term, ==, 0);
+	test_server_run(second);
+
+	return MUNIT_OK;
+}
+
+#endif
diff --git a/test/lib/client.h b/test/lib/client.h
index daab3affa..a42bd8ed6 100644
--- a/test/lib/client.h
+++ b/test/lib/client.h
@@ -140,6 +140,18 @@
 		munit_assert_int(rv_, ==, 0);                            \
 	}
 
+#define EXEC_PARAMS(STMT_ID, LAST_INSERT_ID, ROWS_AFFECTED, ...) \
+	{ \
+		int rv_; \
+		struct value vals_[] = {__VA_ARGS__}; \
+		size_t len_ = sizeof(vals_) / sizeof(vals_[0]); \
+		rv_ = clientSendExec(f->client, STMT_ID, vals_, len_, NULL); \
+		munit_assert_int(rv_, ==, 0); \
+		rv_ = clientRecvResult(f->client, LAST_INSERT_ID, \
+				       ROWS_AFFECTED, NULL); \
+		munit_assert_int(rv_, ==, 0); \
+	}
+
 #define EXEC_SQL(SQL, LAST_INSERT_ID, ROWS_AFFECTED)                    \
 	{                                                               \
 		int rv_;                                                \
diff --git a/test/lib/server.c b/test/lib/server.c
index 89d4060cb..1bfdb6b98 100644
--- a/test/lib/server.c
+++ b/test/lib/server.c
@@ -58,7 +58,7 @@ void test_server_tear_down(struct test_server *s)
 	test_dir_tear_down(s->dir);
 }
 
-void test_server_start(struct test_server *s, const MunitParameter params[])
+void test_server_prepare(struct test_server *s, const MunitParameter params[])
 {
 	int rv;
 
@@ -128,6 +128,11 @@ void test_server_start(struct test_server *s, const MunitParameter params[])
 			munit_assert_int(rv, ==, 0);
 		}
 	}
+}
+
+void test_server_run(struct test_server *s)
+{
+	int rv;
 
 	rv = dqlite_node_start(s->dqlite);
 	munit_assert_int(rv, ==, 0);
@@ -135,6 +140,12 @@ void test_server_start(struct test_server *s, const MunitParameter params[])
 	test_server_client_connect(s, &s->client);
 }
 
+void test_server_start(struct test_server *s, const MunitParameter params[])
+{
+	test_server_prepare(s, params);
+	test_server_run(s);
+}
+
 struct client_proto *test_server_client(struct test_server *s)
 {
 	return &s->client;
diff --git a/test/lib/server.h b/test/lib/server.h
index 7e0dc111b..8b559108e 100644
--- a/test/lib/server.h
+++ b/test/lib/server.h
@@ -35,7 +35,13 @@ void test_server_setup(struct test_server *s,
 /* Cleanup the test server. */
 void test_server_tear_down(struct test_server *s);
 
-/* Start the test server. */
+/* Set up the test server without running it. */
+void test_server_prepare(struct test_server *s, const MunitParameter params[]);
+
+/* Run the test server after setting it up. */
+void test_server_run(struct test_server *s);
+
+/* Start the test server. Equivalent to test_server_prepare + test_server_run. */
 void test_server_start(struct test_server *s, const MunitParameter params[]);
 
 /* Stop the test server. */