Pickle API: Avoid static Series hack, allow multiple Series (#1633)

* Defer iteration parsing

* Add first attempt of createOwningCopy

* Continue...

* seems to work??

* Reenable the Python tests

* Add pickling for Iteration and Series too

* Add pickle support for Series and Iteration

* Document new internal function

Author: franzpoeschel

include/openPMD/Iteration.hpp

@@ -131,6 +131,8 @@ class Iteration : public Attributable
     friend class WriteIterations;
     friend class SeriesIterator;
     friend class internal::AttributableData;
+    template <typename T>
+    friend T &internal::makeOwning(T &self, Series);
 
 public:
     Iteration(Iteration const &) = default;
@@ -258,6 +260,11 @@ class Iteration : public Attributable
         return *m_iterationData;
     }
 
+    inline std::shared_ptr<Data_t> getShared()
+    {
+        return m_iterationData;
+    }
+
     inline void setData(std::shared_ptr<Data_t> data)
     {
         m_iterationData = std::move(data);

include/openPMD/ParticleSpecies.hpp

@@ -35,6 +35,8 @@ class ParticleSpecies : public Container<Record>
     friend class Container<ParticleSpecies>;
     friend class Container<Record>;
     friend class Iteration;
+    template <typename T>
+    friend T &internal::makeOwning(T &self, Series);
 
 public:
     ParticlePatches particlePatches;
@@ -44,6 +46,13 @@ class ParticleSpecies : public Container<Record>
 
     void read();
     void flush(std::string const &, internal::FlushParams const &) override;
+
+    using Data_t = Container<Record>::ContainerData;
+
+    inline std::shared_ptr<Data_t> getShared()
+    {
+        return m_containerData;
+    }
 };
 
 namespace traits

include/openPMD/RecordComponent.hpp

@@ -25,6 +25,7 @@
 #include "openPMD/auxiliary/ShareRaw.hpp"
 #include "openPMD/auxiliary/TypeTraits.hpp"
 #include "openPMD/auxiliary/UniquePtr.hpp"
+#include "openPMD/backend/Attributable.hpp"
 #include "openPMD/backend/BaseRecordComponent.hpp"
 
 #include <array>
@@ -132,6 +133,8 @@ class RecordComponent : public BaseRecordComponent
     friend class DynamicMemoryView;
     friend class internal::RecordComponentData;
     friend class MeshRecordComponent;
+    template <typename T>
+    friend T &internal::makeOwning(T &self, Series);
 
 public:
     enum class Allocation
@@ -523,6 +526,11 @@ OPENPMD_protected
         return *m_recordComponentData;
     }
 
+    inline std::shared_ptr<Data_t> getShared()
+    {
+        return m_recordComponentData;
+    }
+
     inline void setData(std::shared_ptr<internal::RecordComponentData> data)
     {
         m_recordComponentData = std::move(data);

include/openPMD/backend/Attributable.hpp

@@ -124,6 +124,21 @@ namespace internal
     class BaseRecordData;
 
     class RecordComponentData;
+
+    /*
+     * Internal function to turn a handle into an owning handle that will keep
+     * not only itself, but the entire Series alive. Works by hiding a copy of
+     * the Series into the destructor lambda of the internal shared pointer. The
+     * returned handle is entirely safe to use in just the same ways as a normal
+     * handle, just the surrounding Series needs not be kept alive any more
+     * since it is stored within the handle. By storing the Series in the
+     * handle, not in the actual data, reference cycles are avoided.
+     *
+     * Instantiations for T exist for types RecordComponent,
+     * MeshRecordComponent, Mesh, Record, ParticleSpecies, Iteration.
+     */
+    template <typename T>
+    T &makeOwning(T &self, Series);
 } // namespace internal
 
 namespace debug
@@ -157,6 +172,8 @@ class Attributable
     friend class WriteIterations;
     friend class internal::RecordComponentData;
     friend void debug::printDirty(Series const &);
+    template <typename T>
+    friend T &internal::makeOwning(T &self, Series);
 
 protected:
     // tag for internal constructor

include/openPMD/backend/BaseRecord.hpp

@@ -237,6 +237,8 @@ class BaseRecord
     friend class internal::BaseRecordData;
     template <typename, typename, typename>
     friend class internal::ScalarIterator;
+    template <typename T>
+    friend T &internal::makeOwning(T &self, Series);
 
     using Data_t =
         internal::BaseRecordData<T_elem, typename T_RecordComponent::Data_t>;
@@ -256,6 +258,11 @@ class BaseRecord
         return *m_baseRecordData;
     }
 
+    inline std::shared_ptr<Data_t> getShared()
+    {
+        return m_baseRecordData;
+    }
+
     BaseRecord();
 
 protected:

include/openPMD/binding/python/Pickle.hpp

@@ -67,12 +67,9 @@ add_pickle(pybind11::class_<T_Args...> &cl, T_SeriesAccessor &&seriesAccessor)
             std::vector<std::string> const group =
                 t[1].cast<std::vector<std::string> >();
 
-            // Create a new openPMD Series and keep it alive.
-            // This is a big hack for now, but it works for our use
-            // case, which is spinning up remote serial read series
-            // for DASK.
-            static auto series = openPMD::Series(filename, Access::READ_ONLY);
-            return seriesAccessor(series, group);
+            openPMD::Series series(
+                filename, Access::READ_ONLY, "defer_iteration_parsing = true");
+            return seriesAccessor(std::move(series), group);
         }));
 }
 } // namespace openPMD

src/backend/Attributable.cpp

@@ -20,9 +20,12 @@
  */
 #include "openPMD/backend/Attributable.hpp"
 #include "openPMD/Iteration.hpp"
+#include "openPMD/ParticleSpecies.hpp"
+#include "openPMD/RecordComponent.hpp"
 #include "openPMD/Series.hpp"
 #include "openPMD/auxiliary/DerefDynamicCast.hpp"
 #include "openPMD/auxiliary/StringManip.hpp"
+#include "openPMD/backend/Attribute.hpp"
 
 #include <algorithm>
 #include <complex>
@@ -505,4 +508,53 @@ void Attributable::linkHierarchy(Writable &w)
     writable().parent = &w;
     setDirty(true);
 }
+
+namespace internal
+{
+    template <typename T>
+    T &makeOwning(T &self, Series s)
+    {
+        /*
+         * `self` is a handle object such as RecordComponent or Mesh (see
+         * instantiations below).
+         * These objects don't normally keep alive the Series, i.e. as soon as
+         * the Series is destroyed, the handle becomes invalid.
+         * This function modifies the handle such that it actually keeps the
+         * Series alive and behaves otherwise identically.
+         * First, get the internal shared pointer of the handle.
+         */
+        std::shared_ptr<typename T::Data_t> data_ptr = self.T::getShared();
+        auto raw_ptr = data_ptr.get();
+        /*
+         * Now, create a new shared pointer pointing to the same address as the
+         * actual pointer and replace the old internal shared pointer by the new
+         * one.
+         */
+        self.setData(std::shared_ptr<typename T::Data_t>{
+            raw_ptr,
+            /*
+             * Here comes the main trick.
+             * The new shared pointer stores (and thus keeps alive) two items
+             * via lambda capture in its destructor:
+             * 1. The old shared pointer.
+             * 2. The Series.
+             * It's important to notice that these two items are only stored
+             * within the newly created handle, and not internally within the
+             * actual openPMD object model. This means that no reference cycles
+             * can occur.
+             */
+            [s_lambda = std::move(s),
+             data_ptr_lambda = std::move(data_ptr)](auto const *) {
+                /* no-op, the lambda captures simply go out of scope */
+            }});
+        return self;
+    }
+
+    template RecordComponent &makeOwning(RecordComponent &, Series);
+    template MeshRecordComponent &makeOwning(MeshRecordComponent &, Series);
+    template Mesh &makeOwning(Mesh &, Series);
+    template Record &makeOwning(Record &, Series);
+    template ParticleSpecies &makeOwning(ParticleSpecies &, Series);
+    template Iteration &makeOwning(Iteration &, Series);
+} // namespace internal
 } // namespace openPMD

src/binding/python/Iteration.cpp

@@ -23,6 +23,7 @@
 #include "openPMD/backend/Attributable.hpp"
 #include "openPMD/binding/python/Common.hpp"
 #include "openPMD/binding/python/Container.H"
+#include "openPMD/binding/python/Pickle.hpp"
 
 #include <ios>
 #include <sstream>
@@ -33,6 +34,13 @@ void init_Iteration(py::module &m)
     auto py_it_cont = declare_container<PyIterationContainer, Attributable>(
         m, "Iteration_Container");
 
+    // `clang-format on/off` doesn't help here.
+    // Writing this without a macro would lead to a huge diff due to
+    // clang-format.
+#define OPENPMD_AVOID_CLANG_FORMAT auto cl =
+    OPENPMD_AVOID_CLANG_FORMAT
+#undef OPENPMD_AVOID_CLANG_FORMAT
+
     py::class_<Iteration, Attributable>(m, "Iteration")
         .def(py::init<Iteration const &>())
 
@@ -99,5 +107,12 @@ void init_Iteration(py::module &m)
             // garbage collection: return value must be freed before Iteration
             py::keep_alive<1, 0>());
 
+    add_pickle(
+        cl, [](openPMD::Series series, std::vector<std::string> const &group) {
+            uint64_t const n_it = std::stoull(group.at(1));
+            auto &res = series.iterations[n_it];
+            return internal::makeOwning(res, std::move(series));
+        });
+
     finalize_container<PyIterationContainer>(py_it_cont);
 }

src/binding/python/Mesh.cpp

@@ -115,9 +115,10 @@ void init_Mesh(py::module &m)
         .def("set_grid_global_offset", &Mesh::setGridGlobalOffset)
         .def("set_grid_unit_SI", &Mesh::setGridUnitSI);
     add_pickle(
-        cl, [](openPMD::Series &series, std::vector<std::string> const &group) {
+        cl, [](openPMD::Series series, std::vector<std::string> const &group) {
             uint64_t const n_it = std::stoull(group.at(1));
-            return series.iterations[n_it].meshes[group.at(3)];
+            auto &res = series.iterations[n_it].open().meshes[group.at(3)];
+            return internal::makeOwning(res, std::move(series));
         });
 
     finalize_container<PyMeshContainer>(py_m_cont);

src/binding/python/MeshRecordComponent.cpp

@@ -82,12 +82,15 @@ void init_MeshRecordComponent(py::module &m)
             "Relative position of the component on an element "
             "(node/cell/voxel) of the mesh");
     add_pickle(
-        cl, [](openPMD::Series &series, std::vector<std::string> const &group) {
+        cl, [](openPMD::Series series, std::vector<std::string> const &group) {
             uint64_t const n_it = std::stoull(group.at(1));
-            return series.iterations[n_it]
-                .meshes[group.at(3)]
-                       [group.size() < 5 ? MeshRecordComponent::SCALAR
-                                         : group.at(4)];
+            auto &res =
+                series.iterations[n_it]
+                    .open()
+                    .meshes[group.at(3)]
+                           [group.size() < 5 ? MeshRecordComponent::SCALAR
+                                             : group.at(4)];
+            return internal::makeOwning(res, std::move(series));
         });
 
     finalize_container<PyMeshRecordComponentContainer>(py_mrc_cnt);