Document-oriented NoSQL Databases: Oracle NoSQL’s Shift in March 2014

I believe Oracle NoSQL extended and expanded from a key/value store to a document-oriented NoSQL database in March 2014. If you concur, then this was a major shift in the NoSQL arena of document-oriented databases.

Document-Oriented NoSQL Databases

In a narrow interpretation, document-oriented NoSQL databases use JSON or BSON as the query interface data representation. In a wider interpretation, any scalar data type and any composite data type (like maps or array) is available at the query interface without those necessarily complying to the JSON encoding or interpretation.

Some of the known document-oriented database are MongoDB, CouchDB and RethinkDB, amongst many others. The site has a categorization and ranking readily available.

In addition to storing documents (aka, complex data structures with scalar, array and map data types), another important criteria is the ability to query based on the document content by e.g. selecting documents where specific properties have specific values or comply to specific complex predicates (in the general case).

Key/Value Stores

Key/value stores (k/v stores) are different from document-oriented databases. Key/value stores persist a value for a given key. Retrieval is based on the key and values cannot be used in query predicates. Some k/v stores have structured keys or composite keys that support querying a range or set of values.

Oracle NoSQL, for example, is supporting structured and composite keys (,

Aerospike is another example of a k/v store (

K/V Stores: Major Shift

Oracle NoSQL extended its data model from a pure k/v data model to a document-oriented data model in March 2014.

The available simple data types are (cited from

  • Double
  • Float
  • Integer
  • Long
  • Java byte array
  • Java String

And the available complex data types are (cited from

  • Array (An array of values, all of the same type)
  • Enum (An enumeration, represented as an array of strings)
  • Fixed Binary (A fixed-sized binary type used to handle binary data where each record is the same size. It uses less storage than an unrestricted binary field, which requires the length to be stored with the data)
  • Map (An unordered map type where all entries are constrained by a single type)
  • Records

In addition, indexes can be created on the ‘value’ part, aka, the documents: And based on indexes, the database can be queried without having to specify a key:

The latter functionality extends the characteristics of the Oracle NoSQL database from k/v store into a document-oriented NoSQL database. Clearly secondary indexes are not yet a very expressive query language, however, they are a major first step towards it.

What does it mean?

The interesting aspect is that Oracle as a major database vendor moved with a separate and dedicated NoSQL product first into the k/v store space, and now additionally into the document-oriented NoSQL space. There are many ways to read this move, of course, but from a technical viewpoint this emphasizes the importance of a structured data model in the NoSQL space. In addition, OracleNoSQL allows a per-document Avro schema, making it a truly per-document schema database.

Aerospike, for example, also supports complex data types and secondary indexes, even though it is also widely seen as a k/v store (e.g.,


Clearly, the document-oriented segment of NoSQL databases gets more and more crowded and the attention of major players in the database space.

What hasn’t been done is a comparison of the data model and query interface expressiveness of the major players in the document-oriented NoSQL database segment. Once done, and in case that reveals that the expressiveness is the same or almost the same, then the systems are exchangeable from a functionality viewpoint and will distinguish themselves ‘only’ in terms of non-functional properties like latency, throughput, transactions, cost, support, and so on.

Document-oriented NoSQL Databases: How many Joins will you have to implement?

One of the continuously debated items in context of NoSQL databases is the join operation. Let’s listen in a bit:

and there can be many more variations found on the topic of joins on various levels of technical depth.

So, do we need joins in context of NoSQL databases? Do we do joins implemented by NoSQL databases? Are joins outdated concepts that we can live without in context of NoSQL databases? In this blog I try to rationalize the overarching question in principle. Some fact finding first:

(Database) Data Models and Database Management Systems

Data models, like the relational model, the document-model, the hierarchical model, key-value model, graph model, object-oriented model, XML model, etc., are implementations of data structures in a given database management system. Data models define possible data types and their construction rules for more complex types.

For example, the implementation of a relational model might restrict values in tables to be scalar. Another implementation might allow a table as a value, supporting NF2 relations. One system might support the document-model strictly following the JSON model, while others add additional data types in addition to what JSON defines. Some systems do support the notion of references, other so not. Each database implements a data model in any variation it likes to.

Schemata and Database Management Systems

A schema is a particular extension of a domain model, implemented in context of a data model. For example, a domain model might be suppliers, parts and their relationship. This can be implemented in a relational model, a document model or a graph model or any other supported data model.

There is no ‘best’ way of definition a schema. For the same domain, different schemata can be defined depending on the skill of the creator, the knowledge of query access patterns, the amount of restrictions that should be supervised by the database management system and other factors.

For example, in a document model, suppliers, parts and their relationships can be modeled as three separate documents, or in two documents (suppliers and their relationship to parts), or one document – and there are many more variations possible, of course.

Joins and Database Management Systems

Some database management systems implement the join operation in their query interface, some do not. For example, Oracle, MySQL and FoundationDB implement joins, MongoDB, Oracle NoSQL and Aerospike do not. So joins are not necessarily restricted to the relational data model.

Joins and Data Access Paths

With the fact finding under our belt, how many joins will you have to implement? In principle, this is a function of the required data access based on a specific schema. Different schemata of the same domain will require a different number of joins.

Let’s look at a few examples in the supplier – parts domain.

Example 1: No join required

The documents are structured like this:

{"supplier": "superQuality",
     {"part_name": "part_lowQual"}, 
     {"part_name": "part_hiQual"}]

The query: “find the names of all parts for a supplier” does not require a join as the data is already structured so that each supplier contains the set of all parts it supplies.

Example 2: One join required

The documents are structured like this:

{"supplier": "superQuality",
 "parts": [1, 2]
{"part_name": "part_lowQual", "part_id": 1}
{"part_name": "part_hiQual", "part_id": 2}

The query: “find the identifiers and names of all parts for a supplier” requires a join as a supplier only has the identifiers of the parts it ships, not their names.

Example 3: Two joins required

The documents are structured like this:

{"supplier": "superQuality", "supplier_id": "S_55"}
{"part_name": "part_lowQual", "part_id": 1}
{"part_name": "part_hiQual", "part_id": 2}
{"part_id": 1, "supplier_id": "S_55"}

The query: “find the identifiers and names of all parts for a supplier” requires two joins, one to find the objects for a supplier that relate the part identifier to the supplier identifier, and a second one to find the corresponding parts.

Analysis of Examples

The examples have shown empirically that the need for joins is not a function of the data model (document-oriented in this case), but a function of the data access, aka, the number of required data relationship traversals in context of a given schema. If the relationship to be traversed matches the way the data is structured as in Example 1, no join is necessary. As soon as the data is structured differently from the required traversal by the query, joins are necessary (Example 2 and 3).

So, as summary, it is fairly easy to avoid joins. If, and only if, you can structure your data (aka, build your schema) in such a way that it conforms structurally to the queries then you can avoid joins completely (Example 1). I am certain that there are special cases out there for which you can accomplish that, but in general, this is not possible. And, even if it is possible in production, as soon as analysts start analyzing the data sets, they will most likely query along different access paths.

Joins at Query Time vs. Joins at Insert/Update/Delete Time

Above examples clarified that joins are a function of the data access paths. Can joins at query time be avoided entirely by creating data access paths in a certain way?

Yes, it is possible, however, it is a basic trade-off between data query and data manipulation time: reducing the computational effort at run-time, and instead increasing it during insert / update / delete operations. In principle, joins at query time can be avoided if for each access path there is an equivalent data structure in place.

Example 4: Schema refactoring

The documents in this example look like:

{"supplier": "superQuality", "supplier_id": "S_55"}
{"part_name": "part_lowQual", "part_id": 1}
{"part_name": "part_hiQual", "part_id": 2}
{"part_id": 1, "supplier_id": "S_55"}
{"shipper": "fastShipper", "shipper_id": "SH_01"}
{"part_id": 2, "shipper_id": "SH_01"}

Supplier supply parts, however, shippers ship not any part, but only specific parts (maybe for safety reasons). There can be several queries against this document set:

  • Find all parts supplied by a supplier with a given name
  • Find all parts shipped by a shipper with a given name
  • Find all suppliers and shippers for a part with a given name

Each of these queries requires at least one join. The documents can be restructured easily to avoid joins altogether:

{"supplier": "superQuality", "supplier_id": "S_55",
 "parts": [
     {"part_name": "part_lowQual", "part_id": 1}
{"shipper": "fastShipper", "shipper_id": "SH_01",
 "parts": [
     {"part_name": "part_hiQual", "part_id": 2}
{"part_name": "part_lowQual", "part_id": 1,
 "suppliers": [
     {"supplier": "superQuality", "supplier_id": "S_55"}
 "shippers": []}
{"part_name": "part_hiQual", "part_id": 2,
 "suppliers": [],
 "shippers": [
     {"shipper": "fastShipper", "shipper_id": "SH_01"}

The idea is clear: structure the data in such a way that a query can be satisfied with a simple selection. And, the consequence is clear, too: data is duplicated, possibly many times. Which means that an insert, update or delete has to know all the locations where to modify the data and has to modify the data consistently (and ideally within a single transaction).

As a side note, this is the situation that normalization tries to address by ensuring that each data item is only once in the database.

Of course, data duplication will have an impact on the size requirements of main memory an disk space. While there is a change in algorithm complexity, there is also a change in the storage and memory size requirements.

Pre-Joining Data

Pre-joining data allows to avoid joins at query time at the cost of duplicating data at data management time. Alternatively expressed, the implementation of duplication at management time is the cost of avoiding normalization combined with query-time joins.

Is there a way to quantify the effort? In principle, there are as many duplications necessary as joins are to be avoided. This is a rough estimate as many joins are the same except for selection and/or projection specifications. If all joins are abstracted to their join criteria (omitting projection and selection), then this is roughly the amount of duplication required.

The article written by Sarah Mei clearly shows the trade-off between data duplication and joins: She clearly describes many of the issues in context of a specific use case.

“Wait a minute, I don’t have joins and it works anyway!”

But, where are the joins? NoSQL databases that do not implement the join operator in their query interface are in use and production.

If not expressed as query, joins are found either in the application system logic or the interface logic, depending on the design. Most likely these are nested-loop joins or hash-based joins (less likely) or a series of selections with the application logic combining the intermediary query results into the final result data set.

And they are not joins on the complete data set either, but usually have some selection criteria. So the application system logic roughly corresponds to the optimized operator tree of a database query sub-system and in all actuality there might be many joins implemented that way throughout the application logic.

The joins are in fact implemented, just not by using a join operator on the database interface, but inside the application logic. This means that the database cannot optimize the execution, plus there are several queries coming from the application logic putting load on the database system.

And this opens up yet another trade-off: data duplication vs. application logic complexity. If the data is structured in such a way that joins are avoided (at the cost of duplication), then the application logic complexity will be reduced also (from algorithms implementing joins to algorithms issuing queries with selections/projections).

Of course, while the application logic complexity is reduced, the data management logic complexity increased as it has to manage duplicate data consistently across the database.

Summary: Are joins required? Yes. Are joins implemented? Yes.

In my mind there is no question that joins are in general needed and actually implemented today, even if the database does not support a join operator directly and even if there are opinions that joins are not needed. I don’t really understand why there is a discussion about this in the first place as the need for a join is a function of the data schema, not the data model.

The fact that a relational database has the capability of joins does not mean you must use it. And the fact that a NoSQL database does not support joins at their query interface does not mean joins are not needed.

At the heart an architecture and engineering decision has to be made (implicitly or explicitly) of how many joins are implemented through data duplication and how many joins are implemented through algorithms in the application logic layer (if there is not join operator available at the database query interface).

It’s that easy.


Document Projection (Part 2): Definition

What does projection in context of JSON structures or documents actually mean? What should the definition of “projection” be? There are several possibilities discussed next.

Document Projection: Complete Branch

Projection in JSON is projecting a branch of the JSON data structure tree, not projecting a single object or value. To request a projection, a property (projection) path in dot notation could be used (and actually is in many systems). The result of a projection is a valid JSON document containing the specified branch(es).

An example document is

{"a": {"b": {"c": 3, "d": 4, "e": 5}}}

Let’s go through a series of projections in the following.

  • Projection path: “a.b.c”
  • Result: {“a”: {“b”: {“c”: 3}}}
  • Projection path: “a.b”
  • Result: {“a”: {“b”: {“c”: 3, “d”: 4, “e”: 5}}}
  • Projection path: “a.e”
  • Result: {}

The result contains the full path of the projection (or more, but not less). If the requested projection path does not exist, the result is the empty document as none of its properties matches the projection path. The empty projection path “” is not defined, meaning, a projection must name at least one property, and that will be a top-level property in a document.

Several branches can be projected concurrently.

  • Projection paths: “a.b.c”, “a.b.d”
  • Result: {“a”: {“b”: {“c”: 3, “d”: 4}}}

The resulting document contains the combination of all of the branches that result in a valid projection. Redundant projection path specification is possible if one of the projection paths is a sub-path of another one. However, the result document is the same if redundancy is present or absent.

Document Projection: Partial Branch

It might be possible that the whole projection path does not exist, but a part of it. In this case it is a possibility to add the existing result up to that point (MongoDB follows this approach). This results in partial paths whereby the value of their last property is the empty document.

For example, “a.b.f” would result in {“a”: {“b”: {}}}. “a” and “b” exist in the example document, “f”, however, does not.

In my opinion, while possibly useful in some cases, I would not make this the default or standard definition as a result is returned that is incomplete and I could argue that it is in fact incorrect since the value of “b” is not the empty document (I could envision a configuration setting that provides these partial branches if needed).

Document Projection: Value

Wait a minute, why does the result document have to consist of full paths?

The reason is based on the implicit restriction on JSON documents that there can be only one property of a given name on the same level in a document. “Implicit” because the JSON definition ( does not mandate the restriction, but many implementations do: property names on the same level of embedding have to be unique.

For example:

{"x": {"b": {"c": 3, "d": 4}}, 
 "y": {"e": {"c": 3, "d": 4}}}

is a perfectly valid document where the property names are unique on every level. So let’s get back to projection and let’s for a moment assume that projection actually returns the value at the end of the path, omitting the path to the property value itself. So,

  • Projection path: “x.b.c”
  • Result: {“c”: 3}

So far so good.

  • Projection paths: “x.b.c”, “y.e.c”
  • Result: ?

What should the result be? The basic assumption is that a projection on a document returns a document. But “x.b.c” and “y.e.c” both return {“c”: 3} as separate documents, but not one document.

  • One possible result could be an array with two documents. But arrays are in general not considered valid top level documents (again, the JSON definition would allow that).
  • To mitigate that, the array could be the value of a property: {“result”: [{“c”: 3}, {“c”: 3}]}. But this would conflict with a document that happens to have a property “result” of type array with two same documents in it.
  • Or, the two documents would have to be embedded in a third one with special names, like {“1”: {“c”: 3}, “2”: {“c”: 3}}. But then, the original document does not have the properties “1” or “2”.

Based on this discussion having projection results being full paths is simpler and direct.

Projection – Result Correspondence Principle

There is also another argument from the user’s viewpoint. If a user wants to project “x.b.c”, then the user might want to access the result document after the query returns with “x.b.c” as the access path. From this viewpoint, the path in the query and the path in the result document should match and not require access path transformation.

Array Projection: Complete Access Path

Documents may contain arrays as well as array of arrays, arrays of objects of arrays, etc., in principle any of these composite types can be on any level of the document. Projection therefore has to be defined on arrays also, not just documents.

The principle of project paths is extended to include array index specification. For example, let’s consider this document:

{"a": [{"a1": 1}, {"a2": 2}], 
 "b": {"c": [{"c1": 3}, {"c2": 4}, {"c3": 5}]}, 
 "d": [6, 7]}

Let’s do a few projections (arrays are 0-index based):

  • Projection path: a[0]
  • Result: {“a”: [{“a1”: 1}]}
  • Projection path: b.c[1]
  • Result: {“b”: {“c”: [“c2”: 4]}}
  • Projection paths: a[1], b.c[2].c3
  • Result: {“a”: [{“a2”: 2}], “b”: {“c”: [{“c3”: 5}]}}
  • Projection path: a[7]
  • Result: {}

Like in the case of documents, full paths are requested and full paths are returned, with several paths possible. A projection path referring to a non-existing property will not contribute to the result.

So far, so good, except that the results do not yet conform to the “Projection – Result Correspondence” principle from above: the projection “a[1]” resulted in a correct document, but that result document cannot be accessed with “a[1]” to obtain the value.

Array Projection: Padding

In order to support the “Projection – Result Correspondence” principle array results can be padded with the value “null”. For example:

  • Projection paths: a[1], b.c[2].c3
  • Result: {“a”: [null, {“a2”: 2}], “b”: {“c”: [null, null, {“c3”: 5}]}}

Now it is possible to access the result with “a[1]” or “b.c[2].c3” in order to obtain the proper results. From a user’s perspective this is great as again the paths used to specify the projection can be used to retrieve the values.

Array Projection: Scalar Values

Scalar values in arrays do not pose a specific problem:

  • Projection paths: a[1], d[1], d[2]
  • Result: {“a”: [null, {“a2”: 2}], “d”: [null, 7]}

And their access can be accomplished using the projection paths.


Initially I thought projection is a straight forward function and not worth a discussion in context of document-oriented databases; but then it turned out to be not that clear cut. Nevertheless, the above is a starting point for a strict rationalization of projection in document-oriented databases based on the JSON data model.

Joins: (Almost) Impossible to Avoid in Document-oriented Databases

There is a lot of ‘chatter’ about the concept and support of joins in document-oriented databases. So what is the underling issue?

Joins in RDBMS

‘Join’ in the relational world is an operation on two relations that relates the tuples in these relations with each other based on some comparison criteria on the tuples’ attributes. For example, the comparison can be ‘R1.a = R2.b’ and so for each tuple from the first relation R1 all tuples from the second relation R2 are retrieved and combined that match the comparison, meaning, the attribute ‘a’ must match the attribute ‘b’. A detailed discussion can be found here:

Joins allow to relate data from different relations and the join operator is supported by a relational database management system. A typical use case is to find all parts that a supplier supplies. And, for a given part, find all its suppliers. The suppliers and parts are usually stored in different relations and the data have an m:n relationship with each other.

Joins across Documents?

So why the chatter, then? If a document-oriented database stores data in different document collections and if the documents need to be related to each other, then a join is in order. The example of suppliers and parts applies here in the exact same way.

Now, if a document-oriented database does not support joins, what to do? Well, in reality the join will be performed in some layer above the database in a programming language. If all suppliers have to be displayed for a given part, then a program that computes this result effectively implements a join; it is not done in the database, though.

Pre-joined data in Documents?

Some optimization is possible. If the access pattern follows an 80-20 rule, then document-oriented databases allow some hard-coded optimization. If in 80% of the cases the suppliers for a part are requested, and only in 20% the opposite, then the designer of the document layout could create for each part document a sub-collection ‘supplier’ that contains the suppliers of this part. In 80% of the cases no join is necessary any more as the suppliers are ‘pre-joined’ with the parts they supply, only in the 20% of the cases a join is necessary.

However, this causes what in the relational world is called anomalies: If a supplier is removed, then all part documents have to be searched for this supplier. Or if a supplier is added, then all those part documents have to be updated that are supplied by this supplier. Updating supplier data also requires to search the part documents. Pre-joining is effectively a specific de-normalization activity for performance reasons.

Does the type of relationship matter?

Are there relationships that by their nature can be pre-joined without penalty? A very specific relationship, the part-of relationship, falls into this category. It is a ‘clean’ approach since the life time of the part-of objects are the exact same as the containing object.

Another relationship that feels as if pre-joining makes sense is the 1:1 relationship where two objects are exclusively related to each other. However, this is not really the case as one object would be a property of the other and that then could be done the other way around, too. So the 80-20 rule case applies here, too.

In reality, however, relationship between data are usually a lot more complex then just part-of relationships. This in turn means that joins will be necessary. The only real exception is if the 80-20 rule is really a 100-0 rule. This would mean that all access are the exact same and no joins are necessary.

Underlying Conceptual Foundation

Conceptually as soon as independent entities (i.e. objects in their own right) are related to each other, and if their relationship is traversed in both directions at some point in time during the execution of the application, a join is necessary and factually taking place.

Pre-joining is the materialization of the traversal in one direction. So two pre-joins, one for each direction, are possible. If the pre-join in both directions takes place, no join has to be performed upon retrieval; however, the join functionality was applied at time of update or insert in order to accomplish the pre-joins.

As soon as pre-joins exist, possible update, insert and delete anomalies have to be carefully taken care of as pre-joins are the equivalent to de-normalization and therefore data redundancy. At insert, update and delete time all redundant copies of the objects have to be found and the appropriate functionality applied.

Pre-joins are for read-performance reasons only; they are not a conceptual matter and in fact cause additional work at insert, update or delete time instead; so the computational work shifted, but is not avoided.


‘Join’ is a database operator. The same functionality can be implemented in application code outside or ‘on top’ of the database. Most likely the method or function is not called ‘join’ even though it in fact implements that functionality. So be aware of the situation that a document-oriented database does not implement a join and the engineers claim not to need one. The functionality of a join might just be there under a different name.

Document-oriented Database (NoSQL Database Category)

‘Document-oriented database’ refers to databases that store their data in terms of documents. More specifically, structured documents and, by going with the times,  documents in JSON ( format at that.

Some examples are MongoDB ( or RavenDB (

These databases belong to a category of databases called ‘NoSQL Databases’ ( This category emerged over the last few years and in this blog site is used in that sense.

Databases that do not prescribe the content format or enforce a specific data model can store JSON by definition; so database systems that did not decide on a particular data representation upfront but left it open, can be considered document-oriented databases in case their payloads are documents.