diff --git a/atomic_cell.hh b/atomic_cell.hh
index 9ee671007f..a04fed2edb 100644
--- a/atomic_cell.hh
+++ b/atomic_cell.hh
@@ -6,6 +6,7 @@
 
 #include "bytes.hh"
 #include "timestamp.hh"
+#include "gc_clock.hh"
 #include <cstdint>
 
 template<typename T>
@@ -122,6 +123,9 @@ public:
     ttl_opt ttl() const {
         return atomic_cell::ttl(_data);
     }
+    bytes_view serialize() const {
+        return _data;
+    }
     friend class atomic_cell::one;
 };
 
@@ -171,6 +175,22 @@ public:
     friend class atomic_cell_or_collection;
 };
 
+// Represents a mutation of a collection.  Actual format is determined by collection type,
+// and is:
+//   set:  list of atomic_cell
+//   map:  list of pair<atomic_cell, bytes> (for key/value)
+//   list: tbd, probably ugly
+class collection_mutation {
+public:
+    struct view {
+        bytes_view data;
+    };
+    struct one {
+        bytes data;
+        operator view() const { return { data }; }
+    };
+};
+
 // A variant type that can hold either an atomic_cell, or a serialized collection.
 // Which type is stored is determinied by the schema.
 class atomic_cell_or_collection final {
@@ -178,8 +198,21 @@ class atomic_cell_or_collection final {
 private:
     atomic_cell_or_collection(bytes&& data) : _data(std::move(data)) {}
 public:
+    atomic_cell_or_collection(atomic_cell::one ac) : _data(std::move(ac._data)) {}
     static atomic_cell_or_collection from_atomic_cell(atomic_cell::one data) { return { std::move(data._data) }; }
     atomic_cell::view as_atomic_cell() const { return atomic_cell::view::from_bytes(_data); }
-    // FIXME: insert collection variant here
+    atomic_cell_or_collection(collection_mutation::one cm) : _data(std::move(cm.data)) {}
+    static atomic_cell_or_collection from_collection_mutation(collection_mutation::one data) {
+        return std::move(data.data);
+    }
+    collection_mutation::view as_collection_mutation() const {
+        return collection_mutation::view{_data};
+    }
 };
 
+class column_definition;
+
+int compare_atomic_cell_for_merge(atomic_cell::view left, atomic_cell::view right);
+void merge_column(const column_definition& def,
+        atomic_cell_or_collection& old,
+        const atomic_cell_or_collection& neww);
diff --git a/combine.hh b/combine.hh
new file mode 100644
index 0000000000..71738d2f2b
--- /dev/null
+++ b/combine.hh
@@ -0,0 +1,41 @@
+/*
+ * Copyright (C) 2015 Cloudius Systems, Ltd.
+ */
+
+#pragma once
+
+// combine two sorted uniqued sequences into a single sorted sequence
+// unique elements are copied, duplicate elements are merged with a
+// binary function.
+template <typename InputIterator1,
+          typename InputIterator2,
+          typename OutputIterator,
+          typename Compare,
+          typename Merge>
+OutputIterator
+combine(InputIterator1 begin1, InputIterator1 end1,
+        InputIterator2 begin2, InputIterator2 end2,
+        OutputIterator out,
+        Compare compare,
+        Merge merge) {
+    while (begin1 != end1 && begin2 != end2) {
+        auto& e1 = *begin1;
+        auto& e2 = *begin2;
+        if (compare(e1, e2)) {
+            *out++ = e1;
+            ++begin1;
+        } else if (compare(e2, e1)) {
+            *out++ = e2;
+            ++begin2;
+        } else {
+            *out++ = merge(e1, e2);
+            ++begin1;
+            ++begin2;
+        }
+    }
+    out = std::copy(begin1, end1, out);
+    out = std::copy(begin2, end2, out);
+    return out;
+}
+
+
diff --git a/configure.py b/configure.py
index 306ebe6f35..ca88d094ce 100755
--- a/configure.py
+++ b/configure.py
@@ -255,6 +255,7 @@ urchin_core = (['database.cc',
                  'cql3/abstract_marker.cc',
                  'cql3/cql3.cc',
                  'cql3/cql3_type.cc',
+                 'cql3/operation.cc',
                  'cql3/functions/functions.cc',
                  'cql3/statements/modification_statement.cc',
                  'cql3/statements/update_statement.cc',
diff --git a/cql3/Cql.g b/cql3/Cql.g
index 6797179171..8cb4444ae3 100644
--- a/cql3/Cql.g
+++ b/cql3/Cql.g
@@ -1158,9 +1158,7 @@ columnOperation[operations_type& operations]
 
 columnOperationDifferentiator[operations_type& operations, ::shared_ptr<cql3::column_identifier::raw> key]
     : '=' normalColumnOperation[operations, key]
-#if 0
     | '[' k=term ']' specializedColumnOperation[operations, key, k]
-#endif
     ;
 
 normalColumnOperation[operations_type& operations, ::shared_ptr<cql3::column_identifier::raw> key]
@@ -1196,14 +1194,16 @@ normalColumnOperation[operations_type& operations, ::shared_ptr<cql3::column_ide
 #endif
     ;
 
-#if 0
-specializedColumnOperation[List<Pair<ColumnIdentifier.Raw, Operation.RawUpdate>> operations, ColumnIdentifier.Raw key, Term.Raw k]
+specializedColumnOperation[std::vector<std::pair<shared_ptr<cql3::column_identifier::raw>,
+                                                 shared_ptr<cql3::operation::raw_update>>> operations,
+                           shared_ptr<cql3::column_identifier::raw> key,
+                           shared_ptr<cql3::term::raw> k]
+                           
     : '=' t=term
       {
-          addRawUpdate(operations, key, new Operation.SetElement(k, t));
+          add_raw_update(operations, key, make_shared<cql3::operation::set_element>(k, t));
       }
     ;
-#endif
 
 columnCondition[conditions_type& conditions]
     // Note: we'll reject duplicates later
diff --git a/cql3/maps.hh b/cql3/maps.hh
index 03ea3f27b6..a648540bf0 100644
--- a/cql3/maps.hh
+++ b/cql3/maps.hh
@@ -27,6 +27,8 @@
 
 #include "cql3/abstract_marker.hh"
 #include "cql3/term.hh"
+#include "operation.hh"
+#include "update_parameters.hh"
 
 namespace cql3 {
 
@@ -60,20 +62,20 @@ import org.apache.cassandra.utils.Pair;
  * Static helper methods and classes for maps.
  */
 class maps {
+private:
+    maps() = delete;
 public:
-#if 0
-    private Maps() {}
-
-    public static ColumnSpecification keySpecOf(ColumnSpecification column)
-    {
-        return new ColumnSpecification(column.ksName, column.cfName, new ColumnIdentifier("key(" + column.name + ")", true), ((MapType)column.type).getKeysType());
+    static shared_ptr<column_specification> key_spec_of(column_specification& column) {
+        return ::make_shared<column_specification>(column.ks_name, column.cf_name,
+                ::make_shared<column_identifier>(sprint("key(%s)", *column.name), true),
+                dynamic_pointer_cast<map_type_impl>(column.type)->get_keys_type());
     }
 
-    public static ColumnSpecification valueSpecOf(ColumnSpecification column)
-    {
-        return new ColumnSpecification(column.ksName, column.cfName, new ColumnIdentifier("value(" + column.name + ")", true), ((MapType)column.type).getValuesType());
+    static shared_ptr<column_specification> value_spec_of(column_specification& column) {
+        return ::make_shared<column_specification>(column.ks_name, column.cf_name,
+                ::make_shared<column_identifier>(sprint("value(%s)", *column.name), true),
+                dynamic_pointer_cast<map_type_impl>(column.type)->get_values_type());
     }
-#endif
 
     class literal : public term::raw {
     public:
@@ -329,51 +331,47 @@ public:
             Putter.doPut(t, cf, prefix, column, params);
         }
     }
+#endif
 
-    public static class SetterByKey extends Operation
-    {
-        private final Term k;
-
-        public SetterByKey(ColumnDefinition column, Term k, Term t)
-        {
-            super(column, t);
-            this.k = k;
+    class setter_by_key : public operation {
+        const shared_ptr<term> _k;
+    public:
+        setter_by_key(column_definition& column, shared_ptr<term> k, shared_ptr<term> t)
+            : operation(column, std::move(t)), _k(std::move(k)) {
         }
 
-        @Override
-        public void collectMarkerSpecification(VariableSpecifications boundNames)
-        {
-            super.collectMarkerSpecification(boundNames);
-            k.collectMarkerSpecification(boundNames);
+        virtual void collect_marker_specification(shared_ptr<variable_specifications> bound_names) override {
+            operation::collect_marker_specification(bound_names);
+            _k->collect_marker_specification(bound_names);
         }
 
-        public void execute(ByteBuffer rowKey, ColumnFamily cf, Composite prefix, UpdateParameters params) throws InvalidRequestException
-        {
-            assert column.type.isMultiCell() : "Attempted to set a value for a single key on a frozen map";
-            ByteBuffer key = k.bindAndGet(params.options);
-            ByteBuffer value = t.bindAndGet(params.options);
-            if (key == null)
-                throw new InvalidRequestException("Invalid null map key");
-
-            CellName cellName = cf.getComparator().create(prefix, column, key);
-
-            if (value == null)
-            {
-                cf.addColumn(params.makeTombstone(cellName));
+        virtual void execute(mutation& m, const clustering_prefix& prefix, const update_parameters& params) override {
+            using exceptions::invalid_request_exception;
+            assert(column.type->is_multi_cell()); // "Attempted to set a value for a single key on a frozen map"m
+            bytes_opt key = _k->bind_and_get(params._options);
+            bytes_opt value = _t->bind_and_get(params._options);
+            if (!key) {
+                throw invalid_request_exception("Invalid null map key");
             }
-            else
-            {
-                // We don't support value > 64K because the serialization format encode the length as an unsigned short.
-                if (value.remaining() > FBUtilities.MAX_UNSIGNED_SHORT)
-                    throw new InvalidRequestException(String.format("Map value is too long. Map values are limited to %d bytes but %d bytes value provided",
-                                                                    FBUtilities.MAX_UNSIGNED_SHORT,
-                                                                    value.remaining()));
-
-                cf.addColumn(params.makeColumn(cellName, value));
+            if (value && value->size() >= std::numeric_limits<uint16_t>::max()) {
+                throw invalid_request_exception(
+                        sprint("Map value is too long. Map values are limited to %d bytes but %d bytes value provided",
+                               std::numeric_limits<uint16_t>::max(),
+                               value->size()));
+            }
+            auto avalue = value ? params.make_cell(*value) : params.make_dead_cell();
+            map_type_impl::mutation update = { { std::move(*key), std::move(avalue) } };
+            // should have been verified as map earlier?
+            auto ctype = static_pointer_cast<map_type_impl>(column.type);
+            auto col_mut = ctype->serialize_mutation_form(std::move(update));
+            if (column.is_static()) {
+                m.set_static_cell(column, std::move(col_mut));
+            } else {
+                m.set_clustered_cell(prefix, column, std::move(col_mut));
             }
         }
-    }
-
+    };
+#if 0
     public static class Putter extends Operation
     {
         public Putter(ColumnDefinition column, Term t)
diff --git a/cql3/operation.cc b/cql3/operation.cc
new file mode 100644
index 0000000000..6f826e0920
--- /dev/null
+++ b/cql3/operation.cc
@@ -0,0 +1,65 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you 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.
+ */
+
+/*
+ * Copyright (C) 2015 Cloudius Systems, Ltd.
+ */
+
+#include "operation.hh"
+#include "operation_impl.hh"
+#include "maps.hh"
+
+namespace cql3 {
+
+
+shared_ptr<operation>
+operation::set_element::prepare(const sstring& keyspace, column_definition& receiver) {
+    using exceptions::invalid_request_exception;
+    auto rtype = dynamic_pointer_cast<collection_type_impl>(receiver.type);
+    if (!rtype) {
+        throw invalid_request_exception(sprint("Invalid operation (%s) for non collection column %s", receiver, receiver.name()));
+    } else if (!rtype->is_multi_cell()) {
+        throw invalid_request_exception(sprint("Invalid operation (%s) for frozen collection column %s", receiver, receiver.name()));
+    }
+
+    if (&rtype->_kind == &collection_type_impl::kind::list) {
+        abort();
+        // FIXME:
+#if 0
+            Term idx = selector.prepare(keyspace, Lists.indexSpecOf(receiver));
+            Term lval = value.prepare(keyspace, Lists.valueSpecOf(receiver));
+            return new Lists.SetterByIndex(receiver, idx, lval);
+#endif
+    } else if (&rtype->_kind == &collection_type_impl::kind::set) {
+        throw invalid_request_exception(sprint("Invalid operation (%s) for set column %s", receiver, receiver.name()));
+    } else if (&rtype->_kind == &collection_type_impl::kind::map) {
+        auto key = _selector->prepare(keyspace, maps::key_spec_of(*receiver.column_specification));
+        auto mval = _value->prepare(keyspace, maps::value_spec_of(*receiver.column_specification));
+        return make_shared<maps::setter_by_key>(receiver, key, mval);
+    }
+    abort();
+}
+
+bool
+operation::set_element::is_compatible_with(shared_ptr<raw_update> other) {
+    // TODO: we could check that the other operation is not setting the same element
+    // too (but since the index/key set may be a bind variables we can't always do it at this point)
+    return !dynamic_pointer_cast<set_value>(std::move(other));
+}
+
+}
diff --git a/cql3/operation.hh b/cql3/operation.hh
index 16d5f36cd6..d3698d8520 100644
--- a/cql3/operation.hh
+++ b/cql3/operation.hh
@@ -26,13 +26,16 @@
 #define CQL3_OPERATION_HH
 
 #include "core/shared_ptr.hh"
-
+#include "exceptions/exceptions.hh"
 #include "database.hh"
+#include "term.hh"
 
 #include <experimental/optional>
 
 namespace cql3 {
 
+class update_parameters;
+
 #if 0
 package org.apache.cassandra.cql3;
 
@@ -77,6 +80,8 @@ public:
         , _t{t}
     { }
 
+    virtual ~operation() {}
+
     virtual bool uses_function(const sstring& ks_name, const sstring& function_name) const {
         return _t && _t->uses_function(ks_name, function_name);
     }
@@ -172,54 +177,26 @@ public:
 
     class set_value;
 
+    class set_element : public raw_update {
+        const shared_ptr<term::raw> _selector;
+        const shared_ptr<term::raw> _value;
+    public:
+        set_element(shared_ptr<term::raw> selector, shared_ptr<term::raw> value)
+            : _selector(std::move(selector)), _value(std::move(value)) {
+        }
+
+        virtual shared_ptr<operation> prepare(const sstring& keyspace, column_definition& receiver);
 #if 0
-    public static class SetElement implements RawUpdate
-    {
-        private final Term.Raw selector;
-        private final Term.Raw value;
-
-        public SetElement(Term.Raw selector, Term.Raw value)
-        {
-            this.selector = selector;
-            this.value = value;
-        }
-
-        public Operation prepare(String keyspace, ColumnDefinition receiver) throws InvalidRequestException
-        {
-            if (!(receiver.type instanceof CollectionType))
-                throw new InvalidRequestException(String.format("Invalid operation (%s) for non collection column %s", toString(receiver), receiver.name));
-            else if (!(receiver.type.isMultiCell()))
-                throw new InvalidRequestException(String.format("Invalid operation (%s) for frozen collection column %s", toString(receiver), receiver.name));
-
-            switch (((CollectionType)receiver.type).kind)
-            {
-                case LIST:
-                    Term idx = selector.prepare(keyspace, Lists.indexSpecOf(receiver));
-                    Term lval = value.prepare(keyspace, Lists.valueSpecOf(receiver));
-                    return new Lists.SetterByIndex(receiver, idx, lval);
-                case SET:
-                    throw new InvalidRequestException(String.format("Invalid operation (%s) for set column %s", toString(receiver), receiver.name));
-                case MAP:
-                    Term key = selector.prepare(keyspace, Maps.keySpecOf(receiver));
-                    Term mval = value.prepare(keyspace, Maps.valueSpecOf(receiver));
-                    return new Maps.SetterByKey(receiver, key, mval);
-            }
-            throw new AssertionError();
-        }
-
         protected String toString(ColumnSpecification column)
         {
             return String.format("%s[%s] = %s", column.name, selector, value);
         }
 
-        public boolean isCompatibleWith(RawUpdate other)
-        {
-            // TODO: we could check that the other operation is not setting the same element
-            // too (but since the index/key set may be a bind variables we can't always do it at this point)
-            return !(other instanceof SetValue);
-        }
-    }
+#endif
+        virtual bool is_compatible_with(shared_ptr<raw_update> other) override;
+    };
 
+#if 0
     public static class Addition implements RawUpdate
     {
         private final Term.Raw value;
diff --git a/database.cc b/database.cc
index 5fd0ca07de..f3334d6a8d 100644
--- a/database.cc
+++ b/database.cc
@@ -326,7 +326,6 @@ column_family::apply(const mutation& m) {
 }
 
 // Based on org.apache.cassandra.db.AbstractCell#reconcile()
-static inline
 int
 compare_atomic_cell_for_merge(atomic_cell::view left, atomic_cell::view right) {
     if (left.timestamp() != right.timestamp()) {
@@ -347,15 +346,18 @@ compare_atomic_cell_for_merge(atomic_cell::view left, atomic_cell::view right) {
     }
 }
 
-static inline
-int
-compare_for_merge(const column_definition& def,
-                  const std::pair<column_id, atomic_cell_or_collection>& left,
-                  const std::pair<column_id, atomic_cell_or_collection>& right) {
+void
+merge_column(const column_definition& def,
+             atomic_cell_or_collection& old,
+             const atomic_cell_or_collection& neww) {
     if (def.is_atomic()) {
-        return compare_atomic_cell_for_merge(left.second.as_atomic_cell(), right.second.as_atomic_cell());
+        if (compare_atomic_cell_for_merge(old.as_atomic_cell(), neww.as_atomic_cell()) < 0) {
+            // FIXME: move()?
+            old = neww;
+        }
     } else {
-        fail(unimplemented::cause::COLLECTIONS);
+        auto ct = static_pointer_cast<collection_type_impl>(def.type);
+        old = ct->merge(old.as_collection_mutation(), neww.as_collection_mutation());
     }
 }
 
@@ -367,23 +369,24 @@ mutation_partition::apply(schema_ptr schema, const mutation_partition& p) {
         apply_row_tombstone(schema, entry);
     }
 
-    auto merge_cells = [this, schema] (row& old_row, const row& new_row) {
+    auto merge_cells = [this, schema] (row& old_row, const row& new_row, auto&& find_column_def) {
         for (auto&& new_column : new_row) {
             auto col = new_column.first;
             auto i = old_row.find(col);
             if (i == old_row.end()) {
-                _static_row.emplace_hint(i, new_column);
+                old_row.emplace_hint(i, new_column);
             } else {
                 auto& old_column = *i;
-                auto& def = schema->regular_column_at(col);
-                if (compare_for_merge(def, old_column, new_column) < 0) {
-                    old_column.second = new_column.second;
-                }
+                auto& def = find_column_def(col);
+                merge_column(def, old_column.second, new_column.second);
             }
         }
     };
 
-    merge_cells(_static_row, p._static_row);
+    auto find_static_column_def = [schema] (auto col) -> column_definition& { return schema->static_column_at(col); };
+    auto find_regular_column_def = [schema] (auto col) -> column_definition& { return schema->regular_column_at(col); };
+
+    merge_cells(_static_row, p._static_row, find_static_column_def);
 
     for (auto&& entry : p._rows) {
         auto& key = entry.first;
@@ -392,7 +395,7 @@ mutation_partition::apply(schema_ptr schema, const mutation_partition& p) {
             _rows.emplace_hint(i, entry);
         } else {
             i->second.t.apply(entry.second.t);
-            merge_cells(i->second.cells, entry.second.cells);
+            merge_cells(i->second.cells, entry.second.cells, find_regular_column_def);
         }
     }
 }
diff --git a/database.hh b/database.hh
index 785a35ac21..ac65c77ddf 100644
--- a/database.hh
+++ b/database.hh
@@ -94,27 +94,28 @@ public:
     mutation(mutation&&) = default;
     mutation(const mutation&) = default;
 
-    void set_static_cell(const column_definition& def, atomic_cell::one value) {
-        emplace_or_insert(p.static_row(), def.id, std::move(value));
+    void set_static_cell(const column_definition& def, atomic_cell_or_collection value) {
+        update_column(p.static_row(), def, std::move(value));
     }
 
-    void set_clustered_cell(const clustering_prefix& prefix, const column_definition& def, atomic_cell::one value) {
+    void set_clustered_cell(const clustering_prefix& prefix, const column_definition& def, atomic_cell_or_collection value) {
         auto& row = p.clustered_row(serialize_value(*schema->clustering_key_type, prefix));
-        emplace_or_insert(row, def.id, std::move(value));
+        update_column(row, def, std::move(value));
     }
 
-    void set_clustered_cell(const clustering_key& key, const column_definition& def, atomic_cell::one value) {
+    void set_clustered_cell(const clustering_key& key, const column_definition& def, atomic_cell_or_collection value) {
         auto& row = p.clustered_row(key);
-        emplace_or_insert(row, def.id, std::move(value));
+        update_column(row, def, std::move(value));
     }
 private:
-    static void emplace_or_insert(row& row, column_id id, atomic_cell::one&& value) {
+    static void update_column(row& row, const column_definition& def, atomic_cell_or_collection&& value) {
         // our mutations are not yet immutable
+        auto id = def.id;
         auto i = row.lower_bound(id);
         if (i == row.end() || i->first != id) {
-            row.emplace_hint(i, id, atomic_cell_or_collection::from_atomic_cell(std::move(value)));
+            row.emplace_hint(i, id, std::move(value));
         } else {
-            i->second = atomic_cell_or_collection::from_atomic_cell(std::move(value));
+            merge_column(def, i->second, value);
         }
     }
     friend std::ostream& operator<<(std::ostream& os, const mutation& m);
diff --git a/schema.hh b/schema.hh
index 39f58d2b46..d9d169dbb8 100644
--- a/schema.hh
+++ b/schema.hh
@@ -34,6 +34,9 @@ public:
     bool is_compact_value() const;
     const sstring& name_as_text() const;
     const bytes& name() const;
+    friend std::ostream& operator<<(std::ostream& os, const column_definition& cd) {
+        return os << cd.name_as_text();
+    }
 };
 
 struct thrift_schema {
@@ -121,6 +124,9 @@ public:
     column_definition& regular_column_at(column_id id) {
         return _regular_columns[id];
     }
+    column_definition& static_column_at(column_id id) {
+        return _static_columns[id];
+    }
     bool is_last_partition_key(column_definition& def) {
         return &_partition_key[_partition_key.size() - 1] == &def;
     }
diff --git a/tests/urchin/mutation_test.cc b/tests/urchin/mutation_test.cc
index 2689181d8c..e23b833617 100644
--- a/tests/urchin/mutation_test.cc
+++ b/tests/urchin/mutation_test.cc
@@ -66,3 +66,36 @@ BOOST_AUTO_TEST_CASE(test_row_tombstone_updates) {
     m.p.apply_row_tombstone(s, c_key2, tombstone(1, ttl));
     BOOST_REQUIRE_EQUAL(m.p.tombstone_for_row(s, c_key2), tombstone(1, ttl));
 }
+
+BOOST_AUTO_TEST_CASE(test_map_mutations) {
+    auto my_map_type = map_type_impl::get_instance(int32_type, utf8_type, true);
+    auto s = make_lw_shared(schema(some_keyspace, some_column_family,
+        {{"p1", utf8_type}}, {{"c1", int32_type}}, {}, {{"s1", my_map_type}}, utf8_type));
+    column_family cf(s);
+    partition_key key = to_bytes("key1");
+    auto& column = *s->get_column_definition("s1");
+    map_type_impl::mutation mmut1{{int32_type->decompose(101), make_atomic_cell(utf8_type->decompose(sstring("101")))}};
+    mutation m1(key, s);
+    m1.set_static_cell(column, my_map_type->serialize_mutation_form(mmut1));
+    cf.apply(m1);
+    map_type_impl::mutation mmut2{{int32_type->decompose(102), make_atomic_cell(utf8_type->decompose(sstring("102")))}};
+    mutation m2(key, s);
+    m2.set_static_cell(column, my_map_type->serialize_mutation_form(mmut2));
+    cf.apply(m2);
+    map_type_impl::mutation mmut3{{int32_type->decompose(103), make_atomic_cell(utf8_type->decompose(sstring("103")))}};
+    mutation m3(key, s);
+    m3.set_static_cell(column, my_map_type->serialize_mutation_form(mmut3));
+    cf.apply(m3);
+    map_type_impl::mutation mmut2o{{int32_type->decompose(102), make_atomic_cell(utf8_type->decompose(sstring("102 override")))}};
+    mutation m2o(key, s);
+    m2o.set_static_cell(column, my_map_type->serialize_mutation_form(mmut2o));
+    cf.apply(m2o);
+
+    row& r = cf.find_or_create_partition(key).static_row();
+    auto i = r.find(column.id);
+    BOOST_REQUIRE(i != r.end());
+    auto cell = i->second.as_collection_mutation();
+    auto muts = my_map_type->deserialize_mutation_form(cell.data);
+    BOOST_REQUIRE(muts.size() == 3);
+    // FIXME: more strict tests
+}
diff --git a/types.cc b/types.cc
index cb92f18505..16825987c0 100644
--- a/types.cc
+++ b/types.cc
@@ -7,7 +7,11 @@
 #include "types.hh"
 #include "core/print.hh"
 #include "net/ip.hh"
+#include "database.hh"
+#include "util/serialization.hh"
+#include "combine.hh"
 #include <cmath>
+#include <sstream>
 
 template<typename T>
 struct simple_type_traits {
@@ -541,6 +545,332 @@ struct float_type_impl : floating_type_impl<float> {
 };
 
 
+thread_local logging::logger collection_type_impl::_logger("collection_type_impl");
+const size_t collection_type_impl::max_elements;
+
+const collection_type_impl::kind collection_type_impl::kind::map(
+        [] (shared_ptr<cql3::column_specification> collection, bool is_key) -> shared_ptr<cql3::column_specification> {
+            // FIXME: implement
+            // return isKey ? Maps.keySpecOf(collection) : Maps.valueSpecOf(collection);
+            abort();
+        });
+const collection_type_impl::kind collection_type_impl::kind::set(
+        [] (shared_ptr<cql3::column_specification> collection, bool is_key) -> shared_ptr<cql3::column_specification> {
+            // FIXME: implement
+            // return Sets.valueSpecOf(collection);
+            abort();
+        });
+const collection_type_impl::kind collection_type_impl::kind::list(
+        [] (shared_ptr<cql3::column_specification> collection, bool is_key) -> shared_ptr<cql3::column_specification> {
+            // FIXME: implement
+            // return Lists.valueSpecOf(collection);
+            abort();
+        });
+
+shared_ptr<cql3::column_specification>
+collection_type_impl::kind::make_collection_receiver(shared_ptr<cql3::column_specification> collection, bool is_key) const {
+    return _impl(std::move(collection), is_key);
+}
+
+shared_ptr<cql3::column_specification>
+collection_type_impl::make_collection_receiver(shared_ptr<cql3::column_specification> collection, bool is_key) {
+    return _kind.make_collection_receiver(std::move(collection), is_key);
+}
+
+std::vector<atomic_cell::one>
+collection_type_impl::enforce_limit(std::vector<atomic_cell::one> cells, int version) {
+    assert(is_multi_cell());
+    if (version >= 3 || cells.size() <= max_elements) {
+        return cells;
+    }
+    _logger.error("Detected collection with {} elements, more than the {} limit. Only the first {} elements will be returned to the client. "
+            "Please see http://cassandra.apache.org/doc/cql3/CQL.html#collections for more details.", cells.size(), max_elements, max_elements);
+    cells.erase(cells.begin() + max_elements, cells.end());
+    return cells;
+}
+
+bytes
+collection_type_impl::serialize_for_native_protocol(std::vector<atomic_cell::one> cells, int version) {
+    assert(is_multi_cell());
+    cells = enforce_limit(std::move(cells), version);
+    std::vector<bytes> values = serialized_values(std::move(cells));
+    // FIXME: implement
+    abort();
+    // return CollectionSerializer.pack(values, cells.size(), version);
+}
+
+bool
+collection_type_impl::is_compatible_with(abstract_type& previous) {
+    // FIXME: implement
+    abort();
+}
+
+shared_ptr<cql3::cql3_type>
+collection_type_impl::as_cql3_type() {
+    // FIXME: implement
+    abort();
+}
+
+int read_collection_size(bytes_view& in, int version) {
+    if (version >= 3) {
+        return read_simple<int32_t>(in);
+    } else {
+        return read_simple<uint16_t>(in);
+    }
+}
+
+void write_collection_size(std::ostream& out, int size, int version) {
+    if (version >= 3) {
+        serialize_int32(out, size);
+    } else {
+        serialize_int16(out, uint16_t(size));
+    }
+}
+
+bytes read_collection_value(bytes_view& in, int version) {
+    auto size = version >= 3 ? read_simple<int32_t>(in) : read_simple<uint16_t>(in);
+    auto v = read_simple_bytes(in, size);
+    return bytes(v.begin(), v.end());
+}
+
+void write_collection_value(std::ostream& out, int version, data_type type, const boost::any& value) {
+    // We have to copy here, because we can't guess the size.
+    // FIXME: somehow.
+    std::ostringstream tmp;
+    type->serialize(value, tmp);
+    auto val_bytes = tmp.str();
+    if (version >= 3) {
+        serialize_int32(out, int32_t(val_bytes.size()));
+    } else {
+        serialize_int16(out, uint16_t(val_bytes.size()));
+    }
+    out.rdbuf()->sputn(val_bytes.data(), val_bytes.size());
+}
+
+namespace std {
+
+template <>
+struct hash<pair<data_type, data_type>> : private std::hash<data_type> {
+    size_t operator()(const pair<data_type, data_type>& p) const {
+        // don't simply xor, it will generate the same result for sequential
+        // pointers
+        auto f = hash<data_type>::operator()(p.first);
+        auto s = hash<data_type>::operator()(p.second);
+        return f ^ ((s << 7) | s >> (std::numeric_limits<decltype(s)>::digits - 7));
+    }
+};
+
+}
+
+shared_ptr<map_type_impl>
+map_type_impl::get_instance(data_type keys, data_type values, bool is_multi_cell) {
+    auto& map = is_multi_cell ? _instances : _frozen_instances;
+    auto p = std::make_pair(keys, values);
+    auto i = map.find(p);
+    if (i == map.end()) {
+        auto t = make_shared<map_type_impl>(keys, values, is_multi_cell);
+        i = map.insert(std::make_pair(std::move(p), std::move(t))).first;
+    }
+    return i->second;
+}
+
+map_type_impl::map_type_impl(data_type keys, data_type values, bool is_multi_cell)
+        : collection_type_impl("map<" + keys->name() + ", " + values->name() + ">", kind::map)
+        , _keys(std::move(keys))
+        , _values(std::move(values))
+        , _is_multi_cell(is_multi_cell) {
+}
+
+data_type
+map_type_impl::freeze() {
+    if (_is_multi_cell) {
+        return get_instance(_keys, _values, false);
+    } else {
+        return shared_from_this();
+    }
+}
+
+bool
+map_type_impl::is_compatible_with_frozen(collection_type_impl& previous) {
+    assert(!_is_multi_cell);
+    auto* p = dynamic_cast<map_type_impl*>(&previous);
+    if (!p) {
+        return false;
+    }
+    return _keys->is_compatible_with(*p->_keys)
+            && _values->is_compatible_with(*p->_values);
+}
+
+bool
+map_type_impl::is_value_compatible_with_frozen(collection_type_impl& previous) {
+    assert(!_is_multi_cell);
+    auto* p = dynamic_cast<map_type_impl*>(&previous);
+    if (!p) {
+        return false;
+    }
+    return _keys->is_compatible_with(*p->_keys)
+            && _values->is_value_compatible_with(*p->_values);
+}
+
+bool
+map_type_impl::less(bytes_view o1, bytes_view o2) {
+    return compare_maps(_keys, _values, o1, o2) < 0;
+}
+
+int32_t
+map_type_impl::compare_maps(data_type keys, data_type values, bytes_view o1, bytes_view o2) {
+    if (o1.empty()) {
+        return o2.empty() ? 0 : -1;
+    } else if (o2.empty()) {
+        return 1;
+    }
+    int protocol_version = 3;
+    int size1 = read_collection_size(o1, protocol_version);
+    int size2 = read_collection_size(o2, protocol_version);
+    // FIXME: use std::lexicographical_compare()
+    for (int i = 0; i < std::min(size1, size2); ++i) {
+        auto k1 = read_collection_value(o1, protocol_version);
+        auto k2 = read_collection_value(o2, protocol_version);
+        auto cmp = keys->compare(k1, k2);
+        if (cmp != 0) {
+            return cmp;
+        }
+        auto v1 = read_collection_value(o1, protocol_version);
+        auto v2 = read_collection_value(o2, protocol_version);
+        cmp = values->compare(v1, v2);
+        if (cmp != 0) {
+            return cmp;
+        }
+    }
+    return size1 == size2 ? 0 : (size1 < size2 ? -1 : 1);
+}
+
+void
+map_type_impl::serialize(const boost::any& value, std::ostream& out) {
+    return serialize(value, out, 3);
+}
+
+void
+map_type_impl::serialize(const boost::any& value, std::ostream& out, int protocol_version) {
+    auto& m = boost::any_cast<const native_type&>(value);
+    write_collection_size(out, m.size(), protocol_version);
+    for (auto&& kv : m) {
+        write_collection_value(out, protocol_version, _keys, kv.first);
+        write_collection_value(out, protocol_version, _values, kv.second);
+    }
+}
+
+object_opt
+map_type_impl::deserialize(bytes_view v) {
+    return deserialize(v, 3);
+}
+
+object_opt
+map_type_impl::deserialize(bytes_view in, int protocol_version) {
+    if (in.empty()) {
+        return {};
+    }
+    native_type m;
+    auto size = read_collection_size(in, protocol_version);
+    for (int i = 0; i < size; ++i) {
+        bytes kb = read_collection_value(in, protocol_version);
+        auto k = _keys->deserialize(kb);
+        bytes vb = read_collection_value(in, protocol_version);
+        auto v = _values->deserialize(vb);
+        m.insert(m.end(), std::make_pair(std::move(k), std::move(v)));
+    }
+    return { std::move(m) };
+}
+
+sstring
+map_type_impl::to_string(const bytes& b) {
+    // FIXME:
+    abort();
+}
+
+size_t
+map_type_impl::hash(bytes_view v) {
+    // FIXME:
+    abort();
+}
+
+bytes
+map_type_impl::from_string(sstring_view text) {
+    // FIXME:
+    abort();
+}
+
+std::vector<bytes>
+map_type_impl::serialized_values(std::vector<atomic_cell::one> cells) {
+    // FIXME:
+    abort();
+}
+
+auto map_type_impl::deserialize_mutation_form(bytes_view in) -> mutation {
+    auto nr = read_simple<uint32_t>(in);
+    mutation ret;
+    ret.reserve(nr);
+    for (uint32_t i = 0; i != nr; ++i) {
+        // FIXME: we could probably avoid the need for size
+        auto ksize = read_simple<uint32_t>(in);
+        auto key = read_simple_bytes(in, ksize);
+        auto vsize = read_simple<uint32_t>(in);
+        auto value = atomic_cell::view::from_bytes(read_simple_bytes(in, vsize));
+        ret.emplace_back(key, value);
+    }
+    assert(in.empty());
+    return ret;
+}
+
+collection_mutation::one
+map_type_impl::serialize_mutation_form(mutation mut) {
+    std::ostringstream out;
+    auto write32 = [&out] (uint32_t v) {
+        v = net::hton(v);
+        out.write(reinterpret_cast<char*>(&v), sizeof(v));
+    };
+    auto writeb = [&out, write32] (bytes_view v) {
+        write32(v.size());
+        out.write(v.begin(), v.size());
+    };
+    // FIXME: overflow?
+    write32(mut.size());
+    for (auto&& kv : mut) {
+        auto&& k = kv.first;
+        auto&& v = kv.second;
+        writeb(k);
+        writeb(v.serialize());
+    }
+    auto s = out.str();
+    return collection_mutation::one{bytes(s.data(), s.size())};
+}
+
+collection_mutation::one
+map_type_impl::merge(collection_mutation::view a, collection_mutation::view b) {
+    auto aa = deserialize_mutation_form(a.data);
+    auto bb = deserialize_mutation_form(b.data);
+    mutation merged;
+    merged.reserve(aa.size() + bb.size());
+    using element_type = std::pair<bytes_view, atomic_cell::view>;
+    auto compare = [this] (const element_type& e1, const element_type& e2) {
+        return _keys->less(e1.first, e2.first);
+    };
+    auto merge = [this] (const element_type& e1, const element_type& e2) {
+        // FIXME: use std::max()?
+        return std::make_pair(e1.first, compare_atomic_cell_for_merge(e1.second, e2.second) < 0 ? e1.second : e2.second);
+    };
+    combine(aa.begin(), aa.end(),
+            bb.begin(), bb.end(),
+            std::back_inserter(merged),
+            compare,
+            merge);
+    return serialize_mutation_form(merged);
+}
+
+thread_local std::unordered_map<std::pair<data_type, data_type>, map_type_impl::map_type> map_type_impl::_instances;
+thread_local std::unordered_map<std::pair<data_type, data_type>, map_type_impl::map_type> map_type_impl::_frozen_instances;
+
+
 thread_local shared_ptr<abstract_type> int32_type(make_shared<int32_type_impl>());
 thread_local shared_ptr<abstract_type> long_type(make_shared<long_type_impl>());
 thread_local shared_ptr<abstract_type> ascii_type(make_shared<string_type_impl>("ascii", cql3::native_cql3_type::ascii));
diff --git a/types.hh b/types.hh
index 89b9d32c9d..1c44fad9d5 100644
--- a/types.hh
+++ b/types.hh
@@ -16,10 +16,13 @@
 #include "net/byteorder.hh"
 #include "db_clock.hh"
 #include "bytes.hh"
+#include "log.hh"
+#include "atomic_cell.hh"
 
 namespace cql3 {
 
 class cql3_type;
+class column_specification;
 
 }
 
@@ -48,7 +51,7 @@ inline int32_t compare_unsigned(bytes_view v1, bytes_view v2) {
     return (int32_t) (v1.size() - v2.size());
 }
 
-class abstract_type {
+class abstract_type : public enable_shared_from_this<abstract_type> {
     sstring _name;
 public:
     abstract_type(sstring name) : _name(name) {}
@@ -150,10 +153,94 @@ public:
     virtual bool is_collection() { return false; }
     virtual bool is_multi_cell() { return false; }
     virtual ::shared_ptr<cql3::cql3_type> as_cql3_type() = 0;
+    virtual shared_ptr<abstract_type> freeze() { return shared_from_this(); }
 };
 
 using data_type = shared_ptr<abstract_type>;
 
+class collection_type_impl : public abstract_type {
+    static thread_local logging::logger _logger;
+public:
+    static constexpr const size_t max_elements = 65535;
+
+    class kind {
+        std::function<shared_ptr<cql3::column_specification> (shared_ptr<cql3::column_specification> collection, bool is_key)> _impl;
+    public:
+        kind(std::function<shared_ptr<cql3::column_specification> (shared_ptr<cql3::column_specification> collection, bool is_key)> impl)
+            : _impl(std::move(impl)) {}
+        shared_ptr<cql3::column_specification> make_collection_receiver(shared_ptr<cql3::column_specification> collection, bool is_key) const;
+        static const kind map;
+        static const kind set;
+        static const kind list;
+    };
+
+    const kind& _kind;
+
+protected:
+    explicit collection_type_impl(sstring name, const kind& k)
+            : abstract_type(std::move(name)), _kind(k) {}
+public:
+    virtual data_type name_comparator() = 0;
+    virtual data_type value_comparator() = 0;
+    shared_ptr<cql3::column_specification> make_collection_receiver(shared_ptr<cql3::column_specification> collection, bool is_key);
+    virtual bool is_collection() override { return true; }
+    bool is_map() const { return &_kind == &kind::map; }
+    std::vector<atomic_cell::one> enforce_limit(std::vector<atomic_cell::one>, int version);
+    virtual std::vector<bytes> serialized_values(std::vector<atomic_cell::one> cells) = 0;
+    bytes serialize_for_native_protocol(std::vector<atomic_cell::one> cells, int version);
+    virtual bool is_compatible_with(abstract_type& previous) override;
+    virtual bool is_compatible_with_frozen(collection_type_impl& previous) = 0;
+    virtual bool is_value_compatible_with_frozen(collection_type_impl& previous) = 0;
+    virtual shared_ptr<cql3::cql3_type> as_cql3_type() override;
+    virtual collection_mutation::one merge(collection_mutation::view a, collection_mutation::view b) = 0;
+};
+
+using collection_type = shared_ptr<collection_type_impl>;
+
+class map_type_impl final : public collection_type_impl {
+    using map_type = shared_ptr<map_type_impl>;
+    static thread_local std::unordered_map<std::pair<data_type, data_type>, map_type> _instances;
+    static thread_local std::unordered_map<std::pair<data_type, data_type>, map_type> _frozen_instances;
+    data_type _keys;
+    data_type _values;
+    data_type _key_value_pair_type;
+    bool _is_multi_cell;
+public:
+    // type returned by deserialize() and expected by serialize
+    // does not support mutations/ttl/tombstone - purely for I/O.
+    using native_type = std::vector<std::pair<boost::any, boost::any>>;
+    // representation of a map mutation, key/value pairs, value is a mutation itself
+    using mutation = std::vector<std::pair<bytes_view, atomic_cell::view>>;
+    static shared_ptr<map_type_impl> get_instance(data_type keys, data_type values, bool is_multi_cell);
+    map_type_impl(data_type keys, data_type values, bool is_multi_cell);
+    data_type get_keys_type() const { return _keys; }
+    data_type get_values_type() const { return _values; }
+    virtual data_type name_comparator() override { return _keys; }
+    virtual data_type value_comparator() override { return _values; }
+    virtual bool is_multi_cell() override { return _is_multi_cell; }
+    virtual data_type freeze() override;
+    virtual bool is_compatible_with_frozen(collection_type_impl& previous) override;
+    virtual bool is_value_compatible_with_frozen(collection_type_impl& previous) override;
+    virtual bool less(bytes_view o1, bytes_view o2) override;
+    static int32_t compare_maps(data_type keys_comparator, data_type values_comparator,
+                        bytes_view o1, bytes_view o2);
+    virtual bool is_byte_order_comparable() const override { return false; }
+    virtual void serialize(const boost::any& value, std::ostream& out) override;
+    void serialize(const boost::any& value, std::ostream& out, int protocol_version);
+    virtual object_opt deserialize(bytes_view v) override;
+    object_opt deserialize(bytes_view v, int protocol_version);
+    virtual sstring to_string(const bytes& b) override;
+    virtual size_t hash(bytes_view v) override;
+    virtual bytes from_string(sstring_view text) override;
+    virtual std::vector<bytes> serialized_values(std::vector<atomic_cell::one> cells) override;
+    mutation deserialize_mutation_form(bytes_view in);
+    // FIXME: use iterators?
+    collection_mutation::one serialize_mutation_form(mutation mut);
+    virtual collection_mutation::one merge(collection_mutation::view a, collection_mutation::view b) override;
+};
+
+using map_type = shared_ptr<map_type_impl>;
+
 inline
 size_t hash_value(const shared_ptr<abstract_type>& x) {
     return std::hash<abstract_type*>()(x.get());
@@ -323,6 +410,17 @@ T read_simple_exactly(bytes_view& v) {
     return net::ntoh(*reinterpret_cast<const net::packed<T>*>(p));
 }
 
+inline
+bytes_view
+read_simple_bytes(bytes_view& v, size_t n) {
+    if (v.size() < n) {
+        throw marshal_exception();
+    }
+    bytes_view ret(v.begin(), n);
+    v.remove_prefix(n);
+    return ret;
+}
+
 template<typename T>
 object_opt read_simple_opt(bytes_view& v) {
     if (v.empty()) {