Merge branch 'feature/smart-feedback'

12 files changed, 197 insertions, 97 deletions

diff --git a/NEWS b/NEWS
index c2f88d1..4ff4e60 100644
--- a/NEWS
+++ b/NEWS
@@ -1,6 +1,15 @@
 Current release
 ---------------
 
+What's new in psycopg 2.8.3
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+- Added *interval_status* parameter to
+  `~psycopg2.extras.ReplicationCursor.start_replication()` method and other
+  facilities to send automatic replication keepalives at periodic intervals
+  (:ticket:`#913`).
+
+
 What's new in psycopg 2.8.2
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
diff --git a/doc/src/advanced.rst b/doc/src/advanced.rst
index ff0137d..72e353c 100644
--- a/doc/src/advanced.rst
+++ b/doc/src/advanced.rst
@@ -552,8 +552,7 @@ value greater than zero in ``postgresql.conf`` (these changes require a server
 restart).  Create a database ``psycopg2_test``.
 
 Then run the following code to quickly try the replication support out.  This
-is not production code -- it has no error handling, it sends feedback too
-often, etc. -- and it's only intended as a simple demo of logical
+is not production code -- it's only intended as a simple demo of logical
 replication::
 
   from __future__ import print_function
diff --git a/doc/src/extras.rst b/doc/src/extras.rst
index 5d09039..aa316d2 100644
--- a/doc/src/extras.rst
+++ b/doc/src/extras.rst
@@ -270,7 +270,7 @@ The individual messages in the replication stream are represented by
         Replication slots are a feature of PostgreSQL server starting with
         version 9.4.
 
-    .. method:: start_replication(slot_name=None, slot_type=None, start_lsn=0, timeline=0, options=None, decode=False)
+    .. method:: start_replication(slot_name=None, slot_type=None, start_lsn=0, timeline=0, options=None, decode=False, status_interval=10)
 
         Start replication on the connection.
 
@@ -288,6 +288,7 @@ The individual messages in the replication stream are represented by
                         slot (not allowed with physical replication)
         :param decode: a flag indicating that unicode conversion should be
                        performed on messages received from the server
+        :param status_interval: time between feedback packets sent to the server
 
         If a *slot_name* is specified, the slot must exist on the server and
         its type must match the replication type used.
@@ -328,6 +329,14 @@ The individual messages in the replication stream are represented by
         *This parameter should not be set with physical replication or with
         logical replication plugins that produce binary output.*
 
+        Replication stream should periodically send feedback to the database
+        to prevent disconnect via timeout. Feedback is automatically sent when
+        `read_message()` is called or during run of the `consume_stream()`.
+        To specify the feedback interval use *status_interval* parameter.
+        The value of this parameter must be set to at least 1 second, but
+        it can have a fractional part.
+
+
         This function constructs a |START_REPLICATION|_ command and calls
         `start_replication_expert()` internally.
 
@@ -336,10 +345,13 @@ The individual messages in the replication stream are represented by
         `read_message()` in case of :ref:`asynchronous connection
         <async-support>`.
 
+        .. versionchanged:: 2.8.3
+            added the *status_interval* parameter.
+
         .. |START_REPLICATION| replace:: :sql:`START_REPLICATION`
         .. _START_REPLICATION: https://www.postgresql.org/docs/current/static/protocol-replication.html
 
-    .. method:: start_replication_expert(command, decode=False)
+    .. method:: start_replication_expert(command, decode=False, status_interval=10)
 
         Start replication on the connection using provided
         |START_REPLICATION|_ command.
@@ -348,9 +360,13 @@ The individual messages in the replication stream are represented by
             `~psycopg2.sql.Composable` instance for dynamic generation.
         :param decode: a flag indicating that unicode conversion should be
             performed on messages received from the server.
+        :param status_interval: time between feedback packets sent to the server
 
+        .. versionchanged:: 2.8.3
+            added the *status_interval* parameter.
 
-    .. method:: consume_stream(consume, keepalive_interval=10)
+
+    .. method:: consume_stream(consume, keepalive_interval=None)
 
         :param consume: a callable object with signature :samp:`consume({msg})`
         :param keepalive_interval: interval (in seconds) to send keepalive
@@ -373,14 +389,15 @@ The individual messages in the replication stream are represented by
         `ReplicationMessage` class.  See `read_message()` for details about
         message decoding.
 
-        This method also sends keepalive messages to the server in case there
-        were no new data from the server for the duration of
-        *keepalive_interval* (in seconds).  The value of this parameter must
+        This method also sends feedback messages to the server every
+        *keepalive_interval* (in seconds). The value of this parameter must
         be set to at least 1 second, but it can have a fractional part.
+        If the *keepalive_interval* is not specified, the value of
+        *status_interval* specified in the `start_replication()` or
+        `start_replication_expert()` will be used.
 
-        After processing certain amount of messages the client should send a
-        confirmation message to the server.  This should be done by calling
-        `send_feedback()` method on the corresponding replication cursor.  A
+        The client must confirm every processed message by calling
+        `send_feedback()` method on the corresponding replication cursor. A
         reference to the cursor is provided in the `ReplicationMessage` as an
         attribute.
 
@@ -393,9 +410,7 @@ The individual messages in the replication stream are represented by
 
                 def __call__(self, msg):
                     self.process_message(msg.payload)
-
-                    if self.should_send_feedback(msg):
-                        msg.cursor.send_feedback(flush_lsn=msg.data_start)
+                    msg.cursor.send_feedback(flush_lsn=msg.data_start)
 
             consumer = LogicalStreamConsumer()
             cur.consume_stream(consumer)
@@ -408,12 +423,10 @@ The individual messages in the replication stream are represented by
             retains all the WAL segments that might be needed to stream the
             changes via all of the currently open replication slots.
 
-            On the other hand, it is not recommended to send confirmation
-            after *every* processed message, since that will put an
-            unnecessary load on network and the server.  A possible strategy
-            is to confirm after every COMMIT message.
+        .. versionchanged:: 2.8.3
+            changed the default value of the *keepalive_interval* parameter to `!None`.
 
-    .. method:: send_feedback(write_lsn=0, flush_lsn=0, apply_lsn=0, reply=False)
+    .. method:: send_feedback(write_lsn=0, flush_lsn=0, apply_lsn=0, reply=False, force=False)
 
         :param write_lsn: a LSN position up to which the client has written the data locally
         :param flush_lsn: a LSN position up to which the client has processed the
@@ -423,13 +436,19 @@ The individual messages in the replication stream are represented by
                           has applied the changes (physical replication
                           master-slave protocol only)
         :param reply: request the server to send back a keepalive message immediately
+        :param force: force sending a feedback message regardless of status_interval timeout
 
         Use this method to report to the server that all messages up to a
         certain LSN position have been processed on the client and may be
         discarded on the server.
 
-        This method can also be called with all default parameters' values to
-        just send a keepalive message to the server.
+        If the *reply* or *force* parameters are not set, this method will
+        just update internal structures without sending the feedback message
+        to the server. The library sends feedback message automatically
+        when *status_interval* timeout is reached.
+
+        .. versionchanged:: 2.8.3
+            added the *force* parameter.
 
     Low-level replication cursor methods for :ref:`asynchronous connection
     <async-support>` operation.
@@ -463,9 +482,9 @@ The individual messages in the replication stream are represented by
         corresponding connection to block the process until there is more data
         from the server.
 
-        The server can send keepalive messages to the client periodically.
-        Such messages are silently consumed by this method and are never
-        reported to the caller.
+        Last, but not least, this method sends feedback messages when
+        *status_interval* timeout is reached or when keepalive message with
+        reply request arrived from the server.
 
     .. method:: fileno()
 
@@ -481,6 +500,13 @@ The individual messages in the replication stream are represented by
         communication with the server (a data or keepalive message in either
         direction).
 
+    .. attribute:: feedback_timestamp
+
+        A `~datetime` object representing the timestamp at the moment when
+        the last feedback message sent to the server.
+
+        .. versionadded:: 2.8.3
+
     .. attribute:: wal_end
 
        LSN position of the current end of WAL on the server at the
@@ -496,33 +522,21 @@ The individual messages in the replication stream are represented by
 
       def consume(msg):
           # ...
+          msg.cursor.send_feedback(flush_lsn=msg.data_start)
 
-      keepalive_interval = 10.0
+      status_interval = 10.0
       while True:
           msg = cur.read_message()
           if msg:
               consume(msg)
           else:
               now = datetime.now()
-              timeout = keepalive_interval - (now - cur.io_timestamp).total_seconds()
+              timeout = status_interval - (now - cur.feedback_timestamp).total_seconds()
               try:
                   sel = select([cur], [], [], max(0, timeout))
-                  if not any(sel):
-                      cur.send_feedback()  # timed out, send keepalive message
               except InterruptedError:
                   pass  # recalculate timeout and continue
 
-.. warning::
-
-    The :samp:`consume({msg})` function will only be called when there are new
-    database writes on the server e.g. any DML or DDL statement. Depending on
-    your Postgres cluster configuration this might cause the server to run out
-    of disk space if the writes are too far apart. To prevent this from
-    happening you can use `~ReplicationCursor.wal_end` value to periodically
-    send feedback to the server to notify that your replication client has
-    received and processed all the messages.
-
-
 .. index::
     pair: Cursor; Replication
 
diff --git a/lib/extras.py b/lib/extras.py
index f9d3cf6..3fdaaf0 100644
--- a/lib/extras.py
+++ b/lib/extras.py
@@ -561,7 +561,7 @@ class ReplicationCursor(_replicationCursor):
         self.execute(command)
 
     def start_replication(self, slot_name=None, slot_type=None, start_lsn=0,
-                          timeline=0, options=None, decode=False):
+                          timeline=0, options=None, decode=False, status_interval=10):
         """Start replication stream."""
 
         command = "START_REPLICATION "
@@ -615,7 +615,7 @@ class ReplicationCursor(_replicationCursor):
                 command += "%s %s" % (quote_ident(k, self), _A(str(v)))
             command += ")"
 
-        self.start_replication_expert(command, decode=decode)
+        self.start_replication_expert(command, decode=decode, status_interval=status_interval)
 
     # allows replication cursors to be used in select.select() directly
     def fileno(self):
diff --git a/psycopg/pqpath.c b/psycopg/pqpath.c
index 26667eb..b21f397 100644
--- a/psycopg/pqpath.c
+++ b/psycopg/pqpath.c
@@ -1454,12 +1454,20 @@ pq_read_replication_message(replicationCursorObject *repl, replicationMessageObj
     int64_t send_time;
     PyObject *str = NULL, *result = NULL;
     int ret = -1;
+    struct timeval curr_time, feedback_time;
 
     Dprintf("pq_read_replication_message");
 
     *msg = NULL;
     consumed = 0;
 
+    /* Is it a time to send the next feedback message? */
+    gettimeofday(&curr_time, NULL);
+    timeradd(&repl->last_feedback, &repl->status_interval, &feedback_time);
+    if (timercmp(&curr_time, &feedback_time, >=) && pq_send_replication_feedback(repl, 0) < 0) {
+        goto exit;
+    }
+
 retry:
     len = PQgetCopyData(pgconn, &buffer, 1 /* async */);
 
@@ -1552,6 +1560,7 @@ retry:
         (*msg)->send_time  = send_time;
 
         repl->wal_end = wal_end;
+        repl->last_msg_data_start = data_start;
     }
     else if (buffer[0] == 'k') {
         /* Primary keepalive message: msgtype(1), walEnd(8), sendTime(8), reply(1) */
@@ -1565,6 +1574,12 @@ retry:
         Dprintf("pq_read_replication_message: wal_end="XLOGFMTSTR, XLOGFMTARGS(wal_end));
         repl->wal_end = wal_end;
 
+        /* We can safely forward flush_lsn to the wal_end from the server keepalive message
+         * if we know that the client already processed (confirmed) the last XLogData message */
+        if (repl->flush_lsn >= repl->last_msg_data_start && wal_end > repl->flush_lsn) {
+            repl->flush_lsn = wal_end;
+        }
+
         reply = buffer[hdr];
         if (reply && pq_send_replication_feedback(repl, 0) < 0) {
             goto exit;
@@ -1614,7 +1629,8 @@ pq_send_replication_feedback(replicationCursorObject *repl, int reply_requested)
         pq_raise(conn, curs, NULL);
         return -1;
     }
-    gettimeofday(&repl->last_io, NULL);
+    gettimeofday(&repl->last_feedback, NULL);
+    repl->last_io = repl->last_feedback;
 
     return 0;
 }
@@ -1627,7 +1643,7 @@ pq_send_replication_feedback(replicationCursorObject *repl, int reply_requested)
    manages to send keepalive messages to the server as needed.
 */
 int
-pq_copy_both(replicationCursorObject *repl, PyObject *consume, double keepalive_interval)
+pq_copy_both(replicationCursorObject *repl, PyObject *consume)
 {
     cursorObject *curs = &repl->cur;
     connectionObject *conn = curs->conn;
@@ -1636,7 +1652,7 @@ pq_copy_both(replicationCursorObject *repl, PyObject *consume, double keepalive_
     PyObject *tmp = NULL;
     int fd, sel, ret = -1;
     fd_set fds;
-    struct timeval keep_intr, curr_time, ping_time, timeout;
+    struct timeval curr_time, feedback_time, timeout;
 
     if (!PyCallable_Check(consume)) {
         Dprintf("pq_copy_both: expected callable consume object");
@@ -1645,9 +1661,6 @@ pq_copy_both(replicationCursorObject *repl, PyObject *consume, double keepalive_
 
     CLEARPGRES(curs->pgres);
 
-    keep_intr.tv_sec  = (int)keepalive_interval;
-    keep_intr.tv_usec = (long)((keepalive_interval - keep_intr.tv_sec)*1.0e6);
-
     while (1) {
         if (pq_read_replication_message(repl, &msg) < 0) {
             goto exit;
@@ -1662,38 +1675,27 @@ pq_copy_both(replicationCursorObject *repl, PyObject *consume, double keepalive_
             FD_ZERO(&fds);
             FD_SET(fd, &fds);
 
-            /* how long can we wait before we need to send a keepalive? */
+            /* how long can we wait before we need to send a feedback? */
             gettimeofday(&curr_time, NULL);
 
-            timeradd(&repl->last_io, &keep_intr, &ping_time);
-            timersub(&ping_time, &curr_time, &timeout);
+            timeradd(&repl->last_feedback, &repl->status_interval, &feedback_time);
+            timersub(&feedback_time, &curr_time, &timeout);
 
             if (timeout.tv_sec >= 0) {
                 Py_BEGIN_ALLOW_THREADS;
                 sel = select(fd + 1, &fds, NULL, NULL, &timeout);
                 Py_END_ALLOW_THREADS;
-            }
-            else {
-                sel = 0; /* we're past target time, pretend select() timed out */
-            }
-
-            if (sel < 0) {
-                if (errno != EINTR) {
-                    PyErr_SetFromErrno(PyExc_OSError);
-                    goto exit;
-                }
-                if (PyErr_CheckSignals()) {
-                    goto exit;
-                }
-                continue;
-            }
 
-            if (sel == 0) {
-                if (pq_send_replication_feedback(repl, 0) < 0) {
-                    goto exit;
+                if (sel < 0) {
+                    if (errno != EINTR) {
+                        PyErr_SetFromErrno(PyExc_OSError);
+                        goto exit;
+                    }
+                    if (PyErr_CheckSignals()) {
+                        goto exit;
+                    }
                 }
             }
-            continue;
         }
         else {
             tmp = PyObject_CallFunctionObjArgs(consume, msg, NULL);
diff --git a/psycopg/pqpath.h b/psycopg/pqpath.h
index 9d73224..eb0578a 100644
--- a/psycopg/pqpath.h
+++ b/psycopg/pqpath.h
@@ -65,8 +65,7 @@ HIDDEN int pq_execute_command_locked(connectionObject *conn, const char *query,
 RAISES HIDDEN void pq_complete_error(connectionObject *conn);
 
 /* replication protocol support */
-HIDDEN int pq_copy_both(replicationCursorObject *repl, PyObject *consumer,
-                        double keepalive_interval);
+HIDDEN int pq_copy_both(replicationCursorObject *repl, PyObject *consumer);
 HIDDEN int pq_read_replication_message(replicationCursorObject *repl,
                                        replicationMessageObject **msg);
 HIDDEN int pq_send_replication_feedback(replicationCursorObject *repl, int reply_requested);
diff --git a/psycopg/replication_cursor.h b/psycopg/replication_cursor.h
index db549c8..1b92b09 100644
--- a/psycopg/replication_cursor.h
+++ b/psycopg/replication_cursor.h
@@ -42,13 +42,16 @@ typedef struct replicationCursorObject {
     int         decode:1;         /* if we should use character decoding on the messages */
 
     struct timeval last_io;       /* timestamp of the last exchange with the server */
-    struct timeval keepalive_interval;   /* interval for keepalive messages in replication mode */
+    struct timeval status_interval;   /* time between status packets sent to the server */
 
     XLogRecPtr  write_lsn;        /* LSNs for replication feedback messages */
     XLogRecPtr  flush_lsn;
     XLogRecPtr  apply_lsn;
 
     XLogRecPtr  wal_end;          /* WAL end pointer from the last exchange with the server */
+
+    XLogRecPtr  last_msg_data_start; /* WAL pointer to the last non-keepalive message from the server */
+    struct timeval last_feedback; /* timestamp of the last feedback message to the server */
 } replicationCursorObject;
 
 
diff --git a/psycopg/replication_cursor_type.c b/psycopg/replication_cursor_type.c
index 4522076..5fdeaf0 100644
--- a/psycopg/replication_cursor_type.c
+++ b/psycopg/replication_cursor_type.c
@@ -38,8 +38,14 @@
 #include "datetime.h"
 
 
+static void set_status_interval(replicationCursorObject *self, double status_interval)
+{
+    self->status_interval.tv_sec  = (int)status_interval;
+    self->status_interval.tv_usec = (long)((status_interval - self->status_interval.tv_sec)*1.0e6);
+}
+
 #define start_replication_expert_doc \
-"start_replication_expert(command, decode=False) -- Start replication with a given command."
+"start_replication_expert(command, decode=False, status_interval=10) -- Start replication with a given command."
 
 static PyObject *
 start_replication_expert(replicationCursorObject *self,
@@ -49,10 +55,12 @@ start_replication_expert(replicationCursorObject *self,
     connectionObject *conn = self->cur.conn;
     PyObject *res = NULL;
     PyObject *command = NULL;
+    double status_interval = 10;
     long int decode = 0;
-    static char *kwlist[] = {"command", "decode", NULL};
+    static char *kwlist[] = {"command", "decode", "status_interval", NULL};
 
-    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|l", kwlist, &command, &decode)) {
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|ld", kwlist,
+                                     &command, &decode, &status_interval)) {
         return NULL;
     }
 
@@ -64,6 +72,11 @@ start_replication_expert(replicationCursorObject *self,
         goto exit;
     }
 
+    if (status_interval < 1.0) {
+        psyco_set_error(ProgrammingError, curs, "status_interval must be >= 1 (sec)");
+        return NULL;
+    }
+
     Dprintf("start_replication_expert: '%s'; decode: %ld",
         Bytes_AS_STRING(command), decode);
 
@@ -72,6 +85,7 @@ start_replication_expert(replicationCursorObject *self,
         res = Py_None;
         Py_INCREF(res);
 
+        set_status_interval(self, status_interval);
         self->decode = decode;
         gettimeofday(&self->last_io, NULL);
     }
@@ -82,19 +96,19 @@ exit:
 }
 
 #define consume_stream_doc \
-"consume_stream(consumer, keepalive_interval=10) -- Consume replication stream."
+"consume_stream(consumer, keepalive_interval=None) -- Consume replication stream."
 
 static PyObject *
 consume_stream(replicationCursorObject *self,
                                PyObject *args, PyObject *kwargs)
 {
     cursorObject *curs = &self->cur;
-    PyObject *consume = NULL, *res = NULL;
-    double keepalive_interval = 10;
+    PyObject *consume = NULL, *interval = NULL, *res = NULL;
+    double keepalive_interval = 0;
     static char *kwlist[] = {"consume", "keepalive_interval", NULL};
 
-    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|d", kwlist,
-                                     &consume, &keepalive_interval)) {
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|O", kwlist,
+                                     &consume, &interval)) {
         return NULL;
     }
 
@@ -105,9 +119,23 @@ consume_stream(replicationCursorObject *self,
 
     Dprintf("consume_stream");
 
-    if (keepalive_interval < 1.0) {
-        psyco_set_error(ProgrammingError, curs, "keepalive_interval must be >= 1 (sec)");
-        return NULL;
+    if (interval && interval != Py_None) {
+
+        if (PyFloat_Check(interval)) {
+            keepalive_interval = PyFloat_AsDouble(interval);
+        } else if (PyLong_Check(interval)) {
+            keepalive_interval = PyLong_AsDouble(interval);
+        } else if (PyInt_Check(interval)) {
+            keepalive_interval = PyInt_AsLong(interval);
+        } else {
+            psyco_set_error(ProgrammingError, curs, "keepalive_interval must be int or float");
+            return NULL;
+        }
+
+        if (keepalive_interval < 1.0) {
+            psyco_set_error(ProgrammingError, curs, "keepalive_interval must be >= 1 (sec)");
+            return NULL;
+        }
     }
 
     if (self->consuming) {
@@ -124,8 +152,11 @@ consume_stream(replicationCursorObject *self,
     CLEARPGRES(curs->pgres);
 
     self->consuming = 1;
+    if (keepalive_interval > 0) {
+        set_status_interval(self, keepalive_interval);
+    }
 
-    if (pq_copy_both(self, consume, keepalive_interval) >= 0) {
+    if (pq_copy_both(self, consume) >= 0) {
         res = Py_None;
         Py_INCREF(res);
     }
@@ -159,7 +190,7 @@ read_message(replicationCursorObject *self, PyObject *dummy)
 }
 
 #define send_feedback_doc \
-"send_feedback(write_lsn=0, flush_lsn=0, apply_lsn=0, reply=False) -- Try sending a replication feedback message to the server and optionally request a reply."
+"send_feedback(write_lsn=0, flush_lsn=0, apply_lsn=0, reply=False, force=False) -- Update a replication feedback, optionally request a reply or force sending a feedback message regardless of the timeout."
 
 static PyObject *
 send_feedback(replicationCursorObject *self,
@@ -167,13 +198,13 @@ send_feedback(replicationCursorObject *self,
 {
     cursorObject *curs = &self->cur;
     XLogRecPtr write_lsn = 0, flush_lsn = 0, apply_lsn = 0;
-    int reply = 0;
-    static char* kwlist[] = {"write_lsn", "flush_lsn", "apply_lsn", "reply", NULL};
+    int reply = 0, force = 0;
+    static char* kwlist[] = {"write_lsn", "flush_lsn", "apply_lsn", "reply", "force", NULL};
 
     EXC_IF_CURS_CLOSED(curs);
 
-    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|KKKi", kwlist,
-                                     &write_lsn, &flush_lsn, &apply_lsn, &reply)) {
+    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|KKKii", kwlist,
+                                     &write_lsn, &flush_lsn, &apply_lsn, &reply, &force)) {
         return NULL;
     }
 
@@ -186,7 +217,7 @@ send_feedback(replicationCursorObject *self,
     if (apply_lsn > self->apply_lsn)
         self->apply_lsn = apply_lsn;
 
-    if (pq_send_replication_feedback(self, reply) < 0) {
+    if ((force || reply) && pq_send_replication_feedback(self, reply) < 0) {
         return NULL;
     }
 
@@ -228,6 +259,28 @@ repl_curs_get_io_timestamp(replicationCursorObject *self)
     return res;
 }
 
+#define repl_curs_feedback_timestamp_doc \
+"feedback_timestamp -- the timestamp of the latest feedback message sent to the server"
+
+static PyObject *
+repl_curs_get_feedback_timestamp(replicationCursorObject *self)
+{
+    cursorObject *curs = &self->cur;
+    PyObject *tval, *res = NULL;
+    double seconds;
+
+    EXC_IF_CURS_CLOSED(curs);
+
+    seconds = self->last_feedback.tv_sec + self->last_feedback.tv_usec / 1.0e6;
+
+    tval = Py_BuildValue("(d)", seconds);
+    if (tval) {
+        res = PyDateTime_FromTimestamp(tval);
+        Py_DECREF(tval);
+    }
+    return res;
+}
+
 /* object member list */
 
 #define OFFSETOF(x) offsetof(replicationCursorObject, x)
@@ -259,6 +312,9 @@ static struct PyGetSetDef replicationCursorObject_getsets[] = {
     { "io_timestamp",
       (getter)repl_curs_get_io_timestamp, NULL,
       repl_curs_io_timestamp_doc, NULL },
+    { "feedback_timestamp",
+      (getter)repl_curs_get_feedback_timestamp, NULL,
+      repl_curs_feedback_timestamp_doc, NULL },
     {NULL}
 };
 
diff --git a/psycopg/solaris_support.h b/psycopg/solaris_support.h
index 7b99a66..8cfb409 100644
--- a/psycopg/solaris_support.h
+++ b/psycopg/solaris_support.h
@@ -35,6 +35,13 @@
 extern HIDDEN void timeradd(struct timeval *a, struct timeval *b, struct timeval *c);
 extern HIDDEN void timersub(struct timeval *a, struct timeval *b, struct timeval *c);
 #endif
+
+#ifndef timercmp
+#define timercmp(a, b, cmp)          \
+  (((a)->tv_sec == (b)->tv_sec) ?    \
+   ((a)->tv_usec cmp (b)->tv_usec) : \
+   ((a)->tv_sec  cmp (b)->tv_sec))
+#endif
 #endif
 
 #endif /* !defined(PSYCOPG_SOLARIS_SUPPORT_H) */
diff --git a/psycopg/win32_support.h b/psycopg/win32_support.h
index 57b308b..a49d5de 100644
--- a/psycopg/win32_support.h
+++ b/psycopg/win32_support.h
@@ -43,6 +43,13 @@ extern HIDDEN void timeradd(struct timeval *a, struct timeval *b, struct timeval
 #endif
 
 extern HIDDEN void timersub(struct timeval *a, struct timeval *b, struct timeval *c);
+
+#ifndef timercmp
+#define timercmp(a, b, cmp)          \
+  (((a)->tv_sec == (b)->tv_sec) ?    \
+   ((a)->tv_usec cmp (b)->tv_usec) : \
+   ((a)->tv_sec  cmp (b)->tv_sec))
+#endif
 #endif
 
 #endif /* !defined(PSYCOPG_WIN32_SUPPORT_H) */
diff --git a/tests/test_async_keyword.py b/tests/test_async_keyword.py
index d86fa5e..882011c 100755
--- a/tests/test_async_keyword.py
+++ b/tests/test_async_keyword.py
@@ -194,6 +194,7 @@ class AsyncReplicationTest(ReplicationTestCase):
         def consume(msg):
             # just check the methods
             "%s: %s" % (cur.io_timestamp, repr(msg))
+            "%s: %s" % (cur.feedback_timestamp, repr(msg))
 
             self.msg_count += 1
             if self.msg_count > 3:
diff --git a/tests/test_replication.py b/tests/test_replication.py
index 15d47f6..e4de20b 100755
--- a/tests/test_replication.py
+++ b/tests/test_replication.py
@@ -163,18 +163,20 @@ class ReplicationTest(ReplicationTestCase):
         cur = conn.cursor()
 
         self.create_replication_slot(cur, output_plugin='test_decoding')
-
-        # try with invalid options
-        cur.start_replication(
-            slot_name=self.slot, options={'invalid_param': 'value'})
+        self.make_replication_events()
 
         def consume(msg):
-            pass
-        # we don't see the error from the server before we try to read the data
-        self.assertRaises(psycopg2.DataError, cur.consume_stream, consume)
+            raise StopReplication()
+
+        with self.assertRaises(psycopg2.DataError):
+            # try with invalid options
+            cur.start_replication(
+                slot_name=self.slot, options={'invalid_param': 'value'})
+            cur.consume_stream(consume)
 
         # try with correct command
         cur.start_replication(slot_name=self.slot)
+        self.assertRaises(StopReplication, cur.consume_stream, consume)
 
     @skip_before_postgres(9, 4)  # slots require 9.4
     @skip_repl_if_green
@@ -242,6 +244,7 @@ class AsyncReplicationTest(ReplicationTestCase):
         def consume(msg):
             # just check the methods
             "%s: %s" % (cur.io_timestamp, repr(msg))
+            "%s: %s" % (cur.feedback_timestamp, repr(msg))
             "%s: %s" % (cur.wal_end, repr(msg))
 
             self.msg_count += 1