diff --git a/cluster_client.cpp b/cluster_client.cpp
index 8d6601c7..6587be5a 100644
--- a/cluster_client.cpp
+++ b/cluster_client.cpp
@@ -9,10 +9,32 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
     msgr.ringloop = ringloop;
     msgr.repeer_pgs = [this](osd_num_t peer_osd)
     {
-        // peer_osd just connected or dropped connection
         if (msgr.osd_peer_fds.find(peer_osd) != msgr.osd_peer_fds.end())
         {
-            // really connected :)
+            // peer_osd just connected
+            continue_ops();
+        }
+        else if (unsynced_writes.size())
+        {
+            // peer_osd just dropped connection
+            for (auto op: unsynced_writes)
+            {
+                for (auto & part: op->parts)
+                {
+                    if (part.osd_num == peer_osd && part.done)
+                    {
+                        // repeat this operation
+                        part.osd_num = 0;
+                        part.done = false;
+                        assert(!part.sent);
+                        op->done_count--;
+                    }
+                }
+                if (op->done_count < op->parts.size())
+                {
+                    cur_ops.insert(op);
+                }
+            }
             continue_ops();
         }
     };
@@ -30,40 +52,19 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
 
 void cluster_client_t::continue_ops()
 {
-    for (auto op_it = unsent_ops.begin(); op_it != unsent_ops.end(); )
+    if (retry_timeout_id)
     {
-        cluster_op_t *op = *op_it;
-        if (op->needs_reslice && !op->sent_count)
-        {
-            op->parts.clear();
-            op->done_count = 0;
-            op->needs_reslice = false;
-        }
-        if (!op->parts.size())
-        {
-            unsent_ops.erase(op_it++);
-            execute(op);
-            continue;
-        }
-        if (!op->needs_reslice)
-        {
-            for (auto & op_part: op->parts)
-            {
-                if (!op_part.sent && !op_part.done)
-                {
-                    try_send(op, &op_part);
-                }
-            }
-            if (op->sent_count == op->parts.size() - op->done_count)
-            {
-                unsent_ops.erase(op_it++);
-                sent_ops.insert(op);
-            }
-            else
-                op_it++;
-        }
-        else
-            op_it++;
+        tfd->clear_timer(retry_timeout_id);
+        retry_timeout_id = 0;
+    }
+    if (!pg_count)
+    {
+        // Config is not loaded yet
+        return;
+    }
+    for (auto op_it = cur_ops.begin(); op_it != cur_ops.end(); )
+    {
+        continue_rw(*op_it++);
     }
 }
 
@@ -88,33 +89,53 @@ void cluster_client_t::on_load_config_hook(json11::Json::object & config)
     bs_disk_alignment = config["disk_alignment"].uint64_value();
     bs_bitmap_granularity = config["bitmap_granularity"].uint64_value();
     if (!bs_block_size)
-        bs_block_size = DEFAULT_BLOCK_SIZE;
-    if (!bs_disk_alignment)
-        bs_disk_alignment = DEFAULT_DISK_ALIGNMENT;
-    if (!bs_bitmap_granularity)
-        bs_bitmap_granularity = DEFAULT_BITMAP_GRANULARITY;
     {
-        uint32_t block_order;
-        if ((block_order = is_power_of_two(bs_block_size)) >= 64 || bs_block_size < MIN_BLOCK_SIZE || bs_block_size >= MAX_BLOCK_SIZE)
-            throw std::runtime_error("Bad block size");
+        bs_block_size = DEFAULT_BLOCK_SIZE;
+    }
+    if (!bs_disk_alignment)
+    {
+        bs_disk_alignment = DEFAULT_DISK_ALIGNMENT;
+    }
+    if (!bs_bitmap_granularity)
+    {
+        bs_bitmap_granularity = DEFAULT_BITMAP_GRANULARITY;
+    }
+    uint32_t block_order;
+    if ((block_order = is_power_of_two(bs_block_size)) >= 64 || bs_block_size < MIN_BLOCK_SIZE || bs_block_size >= MAX_BLOCK_SIZE)
+    {
+        throw std::runtime_error("Bad block size");
     }
     if (config.find("pg_stripe_size") != config.end())
     {
         pg_stripe_size = config["pg_stripe_size"].uint64_value();
         if (!pg_stripe_size)
+        {
             pg_stripe_size = DEFAULT_PG_STRIPE_SIZE;
+        }
     }
     if (config["immediate_commit"] == "all")
     {
         // Cluster-wide immediate_commit mode
         immediate_commit = true;
     }
+    else if (config.find("client_dirty_limit") != config.end())
+    {
+        client_dirty_limit = config["client_dirty_limit"].uint64_value();
+        if (!client_dirty_limit)
+        {
+            client_dirty_limit = DEFAULT_CLIENT_DIRTY_LIMIT;
+        }
+    }
     msgr.peer_connect_interval = config["peer_connect_interval"].uint64_value();
     if (!msgr.peer_connect_interval)
+    {
         msgr.peer_connect_interval = DEFAULT_PEER_CONNECT_INTERVAL;
+    }
     msgr.peer_connect_timeout = config["peer_connect_timeout"].uint64_value();
     if (!msgr.peer_connect_timeout)
+    {
         msgr.peer_connect_timeout = DEFAULT_PEER_CONNECT_TIMEOUT;
+    }
 }
 
 void cluster_client_t::on_load_pgs_hook(bool success)
@@ -122,7 +143,11 @@ void cluster_client_t::on_load_pgs_hook(bool success)
     if (success)
     {
         pg_count = st_cli.pg_config.size();
-        continue_ops();
+        for (auto op: offline_ops)
+        {
+            execute(op);
+        }
+        offline_ops.clear();
     }
 }
 
@@ -131,16 +156,24 @@ void cluster_client_t::on_change_hook(json11::Json::object & changes)
     if (pg_count != st_cli.pg_config.size())
     {
         // At this point, all operations should be suspended
-        // And they need to be resliced!
-        for (auto op: unsent_ops)
+        // And they have to be resliced!
+        for (auto op: cur_ops)
         {
             op->needs_reslice = true;
         }
-        for (auto op: sent_ops)
+        for (auto op: unsynced_writes)
         {
             op->needs_reslice = true;
         }
         pg_count = st_cli.pg_config.size();
+        if (pg_count)
+        {
+            for (auto op: offline_ops)
+            {
+                execute(op);
+            }
+            offline_ops.clear();
+        }
     }
     continue_ops();
 }
@@ -153,53 +186,176 @@ void cluster_client_t::on_change_osd_state_hook(uint64_t peer_osd)
     }
 }
 
-// FIXME: Implement OSD_OP_SYNC for immediate_commit == false
+/**
+ * How writes are synced when immediate_commit is false
+ *
+ * WRITE:
+ * 1) copy op to the queue
+ * 2) slice op
+ * 3) connect & send all parts
+ * 4) if any of them fail due to disconnected peers, repeat after reconnecting
+ * 5) if any of them fail due to other errors, fail the operation
+ * 6) if PG count changes before all parts are done, wait for all in-progress parts to finish,
+ *    throw all results away, reslice and resubmit op
+ * 7) when all parts are done, try to continue SYNCs
+ *
+ * SYNC:
+ * 1) add SYNC to the queue after writes
+ * 2) try to continue SYNCs
+ *
+ * Continue SYNCs:
+ * 1) check the queue for a SYNC ready for submission. if there is one:
+ * 2) slice it
+ * 3) connect & send all SYNC parts
+ * 4) if any of them fail due to disconnected peers, repeat SYNC after repeating all writes
+ * 5) if any of them fail due to other errors, fail the SYNC operation
+ */
+
 void cluster_client_t::execute(cluster_op_t *op)
 {
-    if (op->opcode == OSD_OP_SYNC && immediate_commit)
-    {
-        // Syncs are not required in the immediate_commit mode
-        op->retval = 0;
-        std::function<void(cluster_op_t*)>(op->callback)(op);
-        return;
-    }
-    if (op->opcode != OSD_OP_READ && op->opcode != OSD_OP_OUT || !op->inode || !op->len ||
-        op->offset % bs_disk_alignment || op->len % bs_disk_alignment)
+    op->retval = 0;
+    if (op->opcode != OSD_OP_SYNC && op->opcode != OSD_OP_READ && op->opcode != OSD_OP_WRITE ||
+        (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_WRITE) && (!op->inode || !op->len ||
+        op->offset % bs_disk_alignment || op->len % bs_disk_alignment))
     {
         op->retval = -EINVAL;
         std::function<void(cluster_op_t*)>(op->callback)(op);
         return;
     }
-    if (!pg_stripe_size)
+    if (!pg_count)
     {
-        // Config is not loaded yet
-        unsent_ops.insert(op);
+        // Config is not loaded yet, retry after connecting to etcd
+        offline_ops.push_back(op);
+        return;
+    }
+    if (op->opcode == OSD_OP_SYNC)
+    {
+        execute_sync(op);
         return;
     }
     if (op->opcode == OSD_OP_WRITE && !immediate_commit)
     {
-        // Copy operation
+        if (next_writes.size() > 0)
+        {
+            next_writes.push_back(op);
+            return;
+        }
+        if (queued_bytes >= client_dirty_limit)
+        {
+            // Push an extra SYNC operation to flush previous writes
+            cluster_op_t *sync_op = new cluster_op_t;
+            sync_op->is_internal = true;
+            sync_op->opcode = OSD_OP_SYNC;
+            sync_op->callback = [](cluster_op_t* sync_op) {};
+            execute_sync(sync_op);
+            next_writes.push_back(op);
+            queued_bytes = 0;
+            return;
+        }
+        queued_bytes += op->len;
+    }
+    cur_ops.insert(op);
+    continue_rw(op);
+}
+
+void cluster_client_t::continue_rw(cluster_op_t *op)
+{
+    if (!pg_count)
+    {
+        return;
+    }
+    if (op->opcode == OSD_OP_WRITE && !immediate_commit && !op->is_internal)
+    {
+        // Save operation for replay when PG goes out of sync
+        // (primary OSD drops our connection in this case)
         cluster_op_t *op_copy = new cluster_op_t();
+        op_copy->is_internal = true;
+        op_copy->orig_op = op;
         op_copy->opcode = op->opcode;
         op_copy->inode = op->inode;
         op_copy->offset = op->offset;
         op_copy->len = op->len;
         op_copy->buf = malloc(op->len);
+        op_copy->callback = [](cluster_op_t* op_copy)
+        {
+            if (op_copy->orig_op)
+            {
+                // Acknowledge write and forget the original pointer
+                op_copy->orig_op->retval = op_copy->retval;
+                std::function<void(cluster_op_t*)>(op_copy->orig_op->callback)(op_copy->orig_op);
+                op_copy->orig_op = NULL;
+            }
+        };
         memcpy(op_copy->buf, op->buf, op->len);
-        unsynced_ops.push_back(op_copy);
-        unsynced_bytes += op->len;
-        if (inmemory_commit)
+        unsynced_writes.push_back(op_copy);
+        cur_ops.erase(op);
+        cur_ops.insert(op_copy);
+        op = op_copy;
+    }
+    if (!op->parts.size())
+    {
+        // Slice the operation into parts
+        slice_rw(op);
+    }
+    if (!op->needs_reslice)
+    {
+        // Send unsent parts, if they're not subject to change
+        for (auto & op_part: op->parts)
         {
-            // Immediately acknowledge write and continue with the copy
-            op->retval = op->len;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
-            op = op_copy;
-        }
-        if (unsynced_bytes >= inmemory_dirty_limit)
-        {
-            // Push an extra SYNC operation
+            if (!op_part.sent && !op_part.done)
+            {
+                try_send(op, &op_part);
+            }
         }
     }
+    if (!op->sent_count)
+    {
+        if (op->done_count >= op->parts.size())
+        {
+            // Finished successfully
+            // Even if the PG count has changed in meanwhile we treat it as success
+            // because if some operations were invalid for the new PG count we'd get errors
+            cur_ops.erase(op);
+            op->retval = op->len;
+            std::function<void(cluster_op_t*)>(op->callback)(op);
+            continue_sync();
+        }
+        else if (op->retval != 0 && op->retval != -EPIPE)
+        {
+            // Fatal error (not -EPIPE)
+            cur_ops.erase(op);
+            if (!immediate_commit && op->opcode == OSD_OP_WRITE)
+            {
+                for (int i = 0; i < unsynced_writes.size(); i++)
+                {
+                    if (unsynced_writes[i] == op)
+                    {
+                        unsynced_writes.erase(unsynced_writes.begin()+i, unsynced_writes.begin()+i+1);
+                        break;
+                    }
+                }
+            }
+            bool del = op->is_internal;
+            std::function<void(cluster_op_t*)>(op->callback)(op);
+            if (del)
+            {
+                delete op;
+            }
+            continue_sync();
+        }
+        else if (op->needs_reslice)
+        {
+            op->parts.clear();
+            op->done_count = 0;
+            op->needs_reslice = false;
+            continue_rw(op);
+        }
+        // else -EPIPE
+    }
+}
+
+void cluster_client_t::slice_rw(cluster_op_t *op)
+{
     // Slice the request into individual object stripe requests
     // Primary OSDs still operate individual stripes, but their size is multiplied by PG minsize in case of EC
     uint64_t pg_block_size = bs_block_size * pg_part_count;
@@ -213,8 +369,8 @@ void cluster_client_t::execute(cluster_op_t *op)
             part_count++;
         }
     }
+    op->retval = 0;
     op->parts.resize(part_count);
-    bool resend = false;
     int i = 0;
     for (uint64_t stripe = first_stripe; stripe <= last_stripe; stripe += pg_block_size)
     {
@@ -231,22 +387,9 @@ void cluster_client_t::execute(cluster_op_t *op)
                 .sent = false,
                 .done = false,
             };
-            if (!try_send(op, &op->parts[i]))
-            {
-                // Part needs to be sent later
-                resend = true;
-            }
             i++;
         }
     }
-    if (resend)
-    {
-        unsent_ops.insert(op);
-    }
-    else
-    {
-        sent_ops.insert(op);
-    }
 }
 
 bool cluster_client_t::try_send(cluster_op_t *op, cluster_op_part_t *part)
@@ -286,7 +429,7 @@ bool cluster_client_t::try_send(cluster_op_t *op, cluster_op_part_t *part)
             {
                 part->op.send_list.push_back(part->buf, part->len);
             }
-            else
+            else /* if (op->opcode == OSD_OP_READ) */
             {
                 part->op.buf = part->buf;
             }
@@ -301,36 +444,163 @@ bool cluster_client_t::try_send(cluster_op_t *op, cluster_op_part_t *part)
     return false;
 }
 
+void cluster_client_t::execute_sync(cluster_op_t *op)
+{
+    if (immediate_commit)
+    {
+        // Syncs are not required in the immediate_commit mode
+        op->retval = 0;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+    }
+    else if (next_writes.size() > 0 || cur_sync != NULL)
+    {
+        // There are some writes postponed to be executed after SYNC
+        // Push this SYNC after them
+        next_writes.push_back(op);
+    }
+    else
+    {
+        cur_sync = op;
+        continue_sync();
+    }
+}
+
+void cluster_client_t::continue_sync()
+{
+    if (!cur_sync || cur_sync->parts.size() > 0)
+    {
+        // Already submitted
+        return;
+    }
+    cur_sync->retval = 0;
+    std::set<osd_num_t> sync_osds;
+    for (auto prev_op: unsynced_writes)
+    {
+        if (prev_op->done_count < prev_op->parts.size())
+        {
+            // Writes not finished yet
+            return;
+        }
+        for (auto & part: prev_op->parts)
+        {
+            if (part.osd_num)
+            {
+                sync_osds.insert(part.osd_num);
+            }
+        }
+    }
+    if (!sync_osds.size())
+    {
+        // No dirty writes
+        finish_sync();
+        return;
+    }
+    // Check that all OSD connections are still alive
+    for (auto sync_osd: sync_osds)
+    {
+        auto peer_it = msgr.osd_peer_fds.find(sync_osd);
+        if (peer_it == msgr.osd_peer_fds.end())
+        {
+            // SYNC is pointless to send to a non connected OSD
+            return;
+        }
+    }
+    // Post sync to affected OSDs
+    cur_sync->parts.resize(sync_osds.size());
+    int i = 0;
+    for (auto sync_osd: sync_osds)
+    {
+        cur_sync->parts[i] = {
+            .parent = cur_sync,
+            .osd_num = sync_osd,
+            .sent = false,
+            .done = false,
+        };
+        send_sync(cur_sync, &cur_sync->parts[i]);
+        i++;
+    }
+}
+
+void cluster_client_t::finish_sync()
+{
+    if (cur_sync->retval == -EPIPE)
+    {
+        // Retry later
+        cur_sync->parts.clear();
+        cur_sync->retval = 0;
+        cur_sync->sent_count = 0;
+        cur_sync->done_count = 0;
+        return;
+    }
+    std::function<void(cluster_op_t*)>(cur_sync->callback)(cur_sync);
+    if (!cur_sync->retval)
+    {
+        for (auto op: unsynced_writes)
+        {
+            if (op->is_internal)
+            {
+                delete op;
+            }
+        }
+        unsynced_writes.clear();
+    }
+    cur_sync = NULL;
+    int i;
+    for (i = 0; i < next_writes.size() && !cur_sync; i++)
+    {
+        execute(next_writes[i]);
+    }
+    next_writes.erase(next_writes.begin(), next_writes.begin()+i);
+}
+
+void cluster_client_t::send_sync(cluster_op_t *op, cluster_op_part_t *part)
+{
+    auto peer_it = msgr.osd_peer_fds.find(part->osd_num);
+    assert(peer_it != msgr.osd_peer_fds.end());
+    part->sent = true;
+    op->sent_count++;
+    part->op = {
+        .op_type = OSD_OP_OUT,
+        .peer_fd = peer_it->second,
+        .req = {
+            .hdr = {
+                .magic = SECONDARY_OSD_OP_MAGIC,
+                .id = op_id++,
+                .opcode = OSD_OP_SYNC,
+            },
+        },
+        .callback = [this, part](osd_op_t *op_part)
+        {
+            handle_op_part(part);
+        },
+    };
+    part->op.send_list.push_back(part->op.req.buf, OSD_PACKET_SIZE);
+    msgr.outbox_push(&part->op);
+}
+
 void cluster_client_t::handle_op_part(cluster_op_part_t *part)
 {
     cluster_op_t *op = part->parent;
     part->sent = false;
     op->sent_count--;
     part->op.buf = NULL;
-    if (part->op.reply.hdr.retval != part->op.req.rw.len)
+    int expected = part->op.req.hdr.opcode == OSD_OP_SYNC ? 0 : part->op.req.rw.len;
+    if (part->op.reply.hdr.retval != expected)
     {
         // Operation failed, retry
         printf(
-            "Operation part failed on OSD %lu: retval=%ld (expected %u), reconnecting\n",
-            part->osd_num, part->op.reply.hdr.retval, part->op.req.rw.len
+            "Operation failed on OSD %lu: retval=%ld (expected %d), dropping connection\n",
+            part->osd_num, part->op.reply.hdr.retval, expected
         );
         msgr.stop_client(part->op.peer_fd);
-        if (op->sent_count == op->parts.size() - op->done_count - 1)
+        if (part->op.reply.hdr.retval && !retry_timeout_id)
         {
-            // Resend later when OSDs come up
-            // FIXME: Check for different types of errors
-            // FIXME: Repeat operations after a small timeout, for the case when OSD is coming up
-            sent_ops.erase(op);
-            unsent_ops.insert(op);
+            retry_timeout_id = tfd->set_timer(up_wait_retry_interval, false, [this](int) { retry_timeout_id = 0; continue_ops(); });
         }
-        if (op->sent_count == 0 && op->needs_reslice)
+        if (!op->retval || op->retval == -EPIPE)
         {
-            // PG count has changed, reslice the operation
-            unsent_ops.erase(op);
-            op->parts.clear();
-            op->done_count = 0;
-            op->needs_reslice = false;
-            execute(op);
+            // Don't overwrite other errors with -EPIPE
+            op->retval = part->op.reply.hdr.retval;
         }
     }
     else
@@ -338,12 +608,17 @@ void cluster_client_t::handle_op_part(cluster_op_part_t *part)
         // OK
         part->done = true;
         op->done_count++;
-        if (op->done_count >= op->parts.size())
+    }
+    if (op->sent_count == 0)
+    {
+        if (op->opcode == OSD_OP_SYNC)
         {
-            // Finished!
-            sent_ops.erase(op);
-            op->retval = op->len;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
+            assert(op == cur_sync);
+            finish_sync();
+        }
+        else
+        {
+            continue_rw(op);
         }
     }
 }
diff --git a/cluster_client.h b/cluster_client.h
index d12c0ae4..de4fa4b3 100644
--- a/cluster_client.h
+++ b/cluster_client.h
@@ -9,6 +9,7 @@
 #define DEFAULT_PG_STRIPE_SIZE 4*1024*1024
 #define DEFAULT_DISK_ALIGNMENT 4096
 #define DEFAULT_BITMAP_GRANULARITY 4096
+#define DEFAULT_CLIENT_DIRTY_LIMIT 32*1024*1024
 
 struct cluster_op_t;
 
@@ -35,6 +36,8 @@ struct cluster_op_t
     void *buf;
     std::function<void(cluster_op_t*)> callback;
 protected:
+    cluster_op_t *orig_op = NULL;
+    bool is_internal = false;
     bool needs_reslice = false;
     int sent_count = 0, done_count = 0;
     std::vector<cluster_op_part_t> parts;
@@ -53,17 +56,24 @@ class cluster_client_t
     uint64_t bs_bitmap_granularity = 0;
     uint64_t pg_count = 0;
     bool immediate_commit = false;
-    bool inmemory_commit = false;
-    uint64_t inmemory_dirty_limit = 32*1024*1024;
+    // FIXME: Implement inmemory_commit mode. Note that it requires to return overlapping reads from memory.
+    uint64_t client_dirty_limit = 0;
     int log_level;
+    int up_wait_retry_interval = 500; // ms
 
     uint64_t op_id = 1;
     etcd_state_client_t st_cli;
     osd_messenger_t msgr;
-    std::set<cluster_op_t*> sent_ops, unsent_ops;
+    // operations currently in progress
+    std::set<cluster_op_t*> cur_ops;
+    int retry_timeout_id = 0;
     // unsynced operations are copied in memory to allow replay when cluster isn't in the immediate_commit mode
-    std::vector<cluster_op_t*> unsynced_ops;
-    uint64_t unsynced_bytes = 0;
+    // unsynced_writes are replayed in any order (because only the SYNC operation guarantees ordering)
+    std::vector<cluster_op_t*> unsynced_writes;
+    cluster_op_t* cur_sync = NULL;
+    std::vector<cluster_op_t*> next_writes;
+    std::vector<cluster_op_t*> offline_ops;
+    uint64_t queued_bytes = 0;
 
 public:
     cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config);
@@ -71,10 +81,16 @@ public:
 
 protected:
     void continue_ops();
-    void on_load_config_hook(json11::Json::object & cfg);
+    void on_load_config_hook(json11::Json::object & config);
     void on_load_pgs_hook(bool success);
     void on_change_hook(json11::Json::object & changes);
     void on_change_osd_state_hook(uint64_t peer_osd);
+    void continue_rw(cluster_op_t *op);
+    void slice_rw(cluster_op_t *op);
     bool try_send(cluster_op_t *op, cluster_op_part_t *part);
+    void execute_sync(cluster_op_t *op);
+    void continue_sync();
+    void finish_sync();
+    void send_sync(cluster_op_t *op, cluster_op_part_t *part);
     void handle_op_part(cluster_op_part_t *part);
 };
diff --git a/messenger.h b/messenger.h
index 9755d9de..723a0dd8 100644
--- a/messenger.h
+++ b/messenger.h
@@ -131,7 +131,7 @@ struct osd_client_t
     std::deque<osd_op_t*> outbox;
     std::map<int, osd_op_t*> sent_ops;
 
-    // PGs dirtied by this client's primary-writes (FIXME to drop the connection)
+    // PGs dirtied by this client's primary-writes
     std::set<pg_num_t> dirty_pgs;
 
     // Write state
diff --git a/osd_peering.cpp b/osd_peering.cpp
index cf16cd53..5e5f1614 100644
--- a/osd_peering.cpp
+++ b/osd_peering.cpp
@@ -133,6 +133,22 @@ void osd_t::start_pg_peering(pg_num_t pg_num)
             it++;
     }
     dirty_pgs.erase(pg.pg_num);
+    // Drop connections of clients who have this PG in dirty_pgs
+    if (immediate_commit != IMMEDIATE_ALL)
+    {
+        std::vector<int> to_stop;
+        for (auto & cp: c_cli.clients)
+        {
+            if (cp.second.dirty_pgs.find(pg_num) != cp.second.dirty_pgs.end())
+            {
+                to_stop.push_back(cp.first);
+            }
+        }
+        for (auto peer_fd: to_stop)
+        {
+            c_cli.stop_client(peer_fd);
+        }
+    }
     // Calculate current write OSD set
     pg.pg_cursize = 0;
     pg.cur_set.resize(pg.target_set.size());