Loading datastore entities from Python project in Go leads to nested structs slices of slices error

I am writing a module in my Google AppEngine project in Go for performance reasons but need to be able to read from some of the entities I have in datastore. I wrote out the Go code to be able to read the entities I built out in Python but I am getting the following error:

datastore: flattening nested structs leads to a slice of slices: field "Messages"

Model Definitions in Python:

class ModelB(ndb.Model):
    msg_id = ndb.StringProperty(indexed=False)
    cat_ids = ndb.StringProperty(repeated=True, indexed=False)
    list_ids = ndb.StringProperty(repeated=True, indexed=False)
    default_list_id_index = ndb.IntegerProperty(indexed=False)

class ModelA(ndb.Model):
    date_join = ndb.DateTimeProperty(auto_now_add=True)
    name = ndb.StringProperty()
    owner_salutation = ndb.StringProperty(indexed=False)
    owner_email_address = ndb.StringProperty()
    logo_url = ndb.StringProperty(indexed=False)
    ...
    messages = ndb.LocalStructuredProperty(ModelB, name='bm', repeated=True)

And in Go:

type ModelB struct {
	MessageID          string   `datastore:"msg_id,noindex"`
	CategoryIDs        []string `datastore:"cat_ids,noindex"`
	ListIDs            []string `datastore:"list_ids,noindex"`
	DefaultListIDIndex int      `datastore:"default_list_id_index,noindex"`
}

type ModelA struct {
	DateJoin          time.Time `datastore:"date_join,"`
	Name              string    `datastore:"name,"`
	OwnerSalutation   string    `datastore:"owner_salutation,noindex"`
	OwnerEmailAddress string    `datastore:"owner_email_address,"`
	LogoURL           string    `datastore:"logo_url,noindex"`
	Messages          []ModelB  `datastore:"bm,"`
}

Is there something I'm doing wrong here? Is just a feature incompatibility between Go vs Python model definitions?

Attempt to Decode ModelB

Re-define ModelA as follows:

import pb "appengine_internal/datastore"
import proto "code.google.com/p/goprotobuf/proto"

type ModelA struct {
	DateJoin          time.Time `datastore:"date_join,"`
	Name              string    `datastore:"name,"`
	OwnerSalutation   string    `datastore:"owner_salutation,noindex"`
	OwnerEmailAddress string    `datastore:"owner_email_address,"`
	LogoURL           string    `datastore:"logo_url,noindex"`
	Messages          []ModelB  `datastore:"-"`
}

// Load is implemented for the PropertyLoaderSaver interface.

func (seller *ModelA) Load(c <-chan datastore.Property) error {
  f := make(chan datastore.Property, 100)
  for p := range c {
    if p.Name == "bm" {
      var val pb.EntityProto
      err := proto.Unmarshal([]byte(p.Value.(string)), &val)
      if err != nil {
        return err
      }
      //TODO: Store result as a new ModelB
    } else {
      f <- p
    }
  }
  close(f)
  return datastore.LoadStruct(seller, f)
}

But I receive the following error:
proto: required field "{Unknown}" not set

The Go datastore package doesn't support two layers of slices like that. You can have []ModelB, as long as ModelB doesn't contain any slices. Or, you can use ModelB in ModelA, and ModelB can have slices in it. But you can't have both []ModelB and ModelB has slices. See the code for the error condition. Your options:

1. don't do it in Go
2. write your own datastore deserializer to handle this case - this is probably hard
3. change your python data structures to satisfy the Go requirements and rewrite your data

I guess if you dig enough you'll find the answer:

First off, when defining the LocalStructuredProperty properties in Python, you need to set keep_keys=True

class ModelB(ndb.Model):
msg_id = ndb.StringProperty(indexed=False)
cat_ids = ndb.StringProperty(repeated=True, indexed=False)
list_ids = ndb.StringProperty(repeated=True, indexed=False)
default_list_id_index = ndb.IntegerProperty(indexed=False)
class ModelA(ndb.Model):
date_join = ndb.DateTimeProperty(auto_now_add=True)
name = ndb.StringProperty()
owner_salutation = ndb.StringProperty(indexed=False)
owner_email_address = ndb.StringProperty()
logo_url = ndb.StringProperty(indexed=False)
...
messages = ndb.LocalStructuredProperty(ModelB, name='bm', repeated=True, keep_keys=True)

A simple redefinition in my code and mapping over my entities doing a put() on each fixed up the representation.

Then in my Go Code:

type ModelB struct {
MessageID          string   `datastore:"msg_id,noindex"`
CategoryIDs        []string `datastore:"cat_ids,noindex"`
ListIDs            []string `datastore:"list_ids,noindex"`
DefaultListIDIndex int      `datastore:"default_list_id_index,noindex"`
}
type ModelA struct {
DateJoin          time.Time `datastore:"date_join,"`
Name              string    `datastore:"name,"`
OwnerSalutation   string    `datastore:"owner_salutation,noindex"`
OwnerEmailAddress string    `datastore:"owner_email_address,"`
LogoURL           string    `datastore:"logo_url,noindex"`
Messages          []ModelB  `datastore:"-"`
}
// Load is implemented for the PropertyLoaderSaver interface.
func (s *ModelA) Load(c <-chan datastore.Property) (err error) {
f := make(chan datastore.Property, 32)
errc := make(chan error, 1)
defer func() {
if err == nil {
err = <-errc
}
}()
go func() {
defer close(f)
for p := range c {
if p.Name == "bm" {
var b ModelB
err := loadLocalStructuredProperty(&b, []byte(p.Value.(string)))
if err != nil {
errc <- err
return
}
s.Messages = append(s.Messages, b)
} else {
f <- p
}
}
errc <- nil
}()
return datastore.LoadStruct(s, f)
}

I had to copy a bunch from the appengine/datastore package as a key function wasn't exported and to simplify the amount of code I needed to copy, I dropped support for Reference types. I opened a ticket on the issue tracker to see if we can get the loadEntity function exported: https://code.google.com/p/googleappengine/issues/detail?id=10426

import (    
"errors"    
"time"    
"appengine"    
"appengine/datastore"        
pb "appengine_internal/datastore"    
proto "code.google.com/p/goprotobuf/proto"    
)    
func loadLocalStructuredProperty(dst interface{}, raw_proto []byte) error {    
var val pb.EntityProto    
err := proto.Unmarshal(raw_proto, &val)    
if err != nil {    
return err    
}    
return loadEntity(dst, &val)    
}
//Copied from appengine/datastore since its not exported
// loadEntity loads an EntityProto into PropertyLoadSaver or struct pointer.
func loadEntity(dst interface{}, src *pb.EntityProto) (err error) {
c := make(chan datastore.Property, 32)
errc := make(chan error, 1)
defer func() {
if err == nil {
err = <-errc
}
}()
go protoToProperties(c, errc, src)
if e, ok := dst.(datastore.PropertyLoadSaver); ok {
return e.Load(c)
}
return datastore.LoadStruct(dst, c)
}
func protoToProperties(dst chan<- datastore.Property, errc chan<- error, src *pb.EntityProto) {
defer close(dst)
props, rawProps := src.Property, src.RawProperty
for {
var (
x       *pb.Property
noIndex bool
)
if len(props) > 0 {
x, props = props[0], props[1:]
} else if len(rawProps) > 0 {
x, rawProps = rawProps[0], rawProps[1:]
noIndex = true
} else {
break
}
var value interface{}
if x.Meaning != nil && *x.Meaning == pb.Property_INDEX_VALUE {
value = indexValue{x.Value}
} else {
var err error
value, err = propValue(x.Value, x.GetMeaning())
if err != nil {
errc <- err
return
}
}
dst <- datastore.Property{
Name:     x.GetName(),
Value:    value,
NoIndex:  noIndex,
Multiple: x.GetMultiple(),
}
}
errc <- nil
}
func fromUnixMicro(t int64) time.Time {
return time.Unix(t/1e6, (t%1e6)*1e3)
}
// propValue returns a Go value that combines the raw PropertyValue with a
// meaning. For example, an Int64Value with GD_WHEN becomes a time.Time.
func propValue(v *pb.PropertyValue, m pb.Property_Meaning) (interface{}, error) {
switch {
case v.Int64Value != nil:
if m == pb.Property_GD_WHEN {
return fromUnixMicro(*v.Int64Value), nil
} else {
return *v.Int64Value, nil
}
case v.BooleanValue != nil:
return *v.BooleanValue, nil
case v.StringValue != nil:
if m == pb.Property_BLOB {
return []byte(*v.StringValue), nil
} else if m == pb.Property_BLOBKEY {
return appengine.BlobKey(*v.StringValue), nil
} else {
return *v.StringValue, nil
}
case v.DoubleValue != nil:
return *v.DoubleValue, nil
case v.Referencevalue != nil:
return nil, errors.New("Not Implemented!")
}
return nil, nil
}
// indexValue is a Property value that is created when entities are loaded from
// an index, such as from a projection query.
//
// Such Property values do not contain all of the metadata required to be
// faithfully represented as a Go value, and are instead represented as an
// opaque indexValue. Load the properties into a concrete struct type (e.g. by
// passing a struct pointer to Iterator.Next) to reconstruct actual Go values
// of type int, string, time.Time, etc.
type indexValue struct {
value *pb.PropertyValue
}

The solution by someone1 works great but I have many millions of entities and didn't want to have to re-put them all (to add the keep_keys=True to the LocalStructuredProperty).

So, I created a cut-down version of EntityProto which removes the dependency on the key & path etc... Simply replace pb.EntityProto with LocalEntityProto and the existing python-written entities should load OK (I'm using a PropertyLoadSaver for the nested struct).

Disclaimer: I'm only using this to read from Go - I haven't tried writing the same entities back to see if they still load in Python.

import pb "google.golang.org/appengine/internal/datastore"
import proto "github.com/golang/protobuf/proto"
type LocalEntityProto struct {
Kind             *pb.EntityProto_Kind `protobuf:"varint,4,opt,name=kind,enum=appengine.EntityProto_Kind" json:"kind,omitempty"`
KindUri          *string              `protobuf:"bytes,5,opt,name=kind_uri" json:"kind_uri,omitempty"`
Property         []*pb.Property       `protobuf:"bytes,14,rep,name=property" json:"property,omitempty"`
RawProperty      []*pb.Property       `protobuf:"bytes,15,rep,name=raw_property" json:"raw_property,omitempty"`
Rank             *int32               `protobuf:"varint,18,opt,name=rank" json:"rank,omitempty"`
XXX_unrecognized []byte               `json:"-"`
}
func (m *LocalEntityProto) Reset()         { *m = LocalEntityProto{} }
func (m *LocalEntityProto) String() string { return proto.CompactTextString(m) }
func (*LocalEntityProto) ProtoMessage()    {}
func (m *LocalEntityProto) GetKind() pb.EntityProto_Kind {
if m != nil && m.Kind != nil {
return *m.Kind
}
return pb.EntityProto_GD_CONTACT
}
func (m *LocalEntityProto) GetKindUri() string {
if m != nil && m.KindUri != nil {
return *m.KindUri
}
return ""
}
func (m *LocalEntityProto) GetProperty() []*pb.Property {
if m != nil {
return m.Property
}
return nil
}
func (m *LocalEntityProto) GetRawProperty() []*pb.Property {
if m != nil {
return m.RawProperty
}
return nil
}
func (m *LocalEntityProto) GetRank() int32 {
if m != nil && m.Rank != nil {
return *m.Rank
}
return 0
}