Skip to content

Latest commit

 

History

History
154 lines (102 loc) · 15.7 KB

result-set-builders.md

File metadata and controls

154 lines (102 loc) · 15.7 KB

RowBuilder and ResultSetBuilder

In Getting Started, it was noted that, by default, execute! and execute-one! return result sets as (vectors of) hash maps with namespace-qualified keys as-is. If your database naturally produces uppercase column names from the JDBC driver, that's what you'll get. If it produces mixed-case names, that's what you'll get.

Note: Some databases do not return the table name in the metadata by default. If you run into this, you might try adding :ResultSetMetaDataOptions "1" to your db-spec (so it is passed as a property to the JDBC driver when you create connections). If your database supports that, it will perform additional work to try to add table names to the result set metadata. It has been reported that Oracle just plain old does not support table names at all in its JDBC drivers.

The default builder for rows and result sets creates qualified keywords that match whatever case the JDBC driver produces. That builder is next.jdbc.result-set/as-maps but there are several options available:

  • as-maps -- table-qualified keywords as-is, the default, e.g., :ADDRESS/ID, :myTable/firstName,
  • as-unqualified-maps -- simple keywords as-is, e.g., :ID, :firstName,
  • as-lower-maps -- table-qualified lower-case keywords, e.g., :address/id, :mytable/firstname,
  • as-unqualified-lower-maps -- simple lower-case keywords, e.g., :id, :firstname,
  • as-arrays -- table-qualified keywords as-is (vector of column names, followed by vectors of row values),
  • as-unqualified-arrays -- simple keywords as-is,
  • as-lower-arrays -- table-qualified lower-case keywords,
  • as-unqualified-lower-arrays -- simple lower-case keywords.

The reason behind the default is to a) be a simple transform, b) produce qualified keys in keeping with Clojure's direction (with clojure.spec etc), and c) not mess with the data. as-arrays is (slightly) faster than as-maps since it produces less data (vectors of values instead of vectors of hash maps), but the lower options will be slightly slower since they include (conditional) logic to convert strings to lower-case. The unqualified options may be slightly faster than their qualified equivalents but make no attempt to keep column names unique if your SQL joins across multiple tables.

Note: This is a deliberate difference from clojure.java.jdbc which would make column names unique by appending numeric suffixes. It was always poor practice to rely on clojure.java.jdbc's renaming behavior and it added quite an overhead to result set building, which is why next.jdbc does not support it -- use explicit column aliasing in your SQL instead if you want unqualified column names!

In addition, the following generic builders can take :label-fn and :qualifier-fn options to control how the label and qualified are processed. The lower variants above are implemented in terms of these, passing a lower-case function for both of those options.

  • as-modified-maps -- table-qualified keywords,
  • as-unqualified-modified-maps -- simple keywords,
  • as-modified-arrays -- table-qualified keywords,
  • as-unqualified-modified-arrays -- simple keywords.

An example builder that naively converts snake_case database table/column names to kebab-case keywords:

(defn as-kebab-maps [rs opts]
  (let [kebab #(str/replace % #"_" "-")]
    (result-set/as-modified-maps rs (assoc opts :qualifier-fn kebab :label-fn kebab))))

However, a version of as-kebab-maps is built-in, as is as-unqualified-kebab-maps, which both use the ->kebab-case function from the camel-snake-kebab library with as-modified-maps and as-unqualified-modified-maps respectively, so you can just use the built-in result-set/as-kebab-maps (or result-set/as-unqualified-kebab-maps) builder as a :builder-fn option instead of writing your own.

Note: next.jdbc/snake-kebab-opts and next.jdbc/unqualified-snake-kebab-opts exist, providing pre-built options hash maps that contain these :builder-fn options, as well as appropriate :table-fn and :column-fn options for the Friendly SQL Functions so those are often the most convenient way to enable snake/kebab case conversions with next.jdbc.

And finally there are two styles of adapters for the existing builders that let you override the default way that columns are read from result sets. The first style takes a column-reader function, which is called with the ResultSet, the ResultSetMetaData, and the column index, and is expected to read the raw column value from the result set and return it. The result is then passed through read-column-by-index (from ReadableColumn, which may be implemented directly via protocol extension or via metadata on the result of the column-reader function):

  • as-maps-adapter -- adapts an existing map builder function with a new column reader,
  • as-arrays-adapter -- adapts an existing array builder function with a new column reader.

The default column-reader function behavior would be:

(defn default-column-reader
  [^ResultSet rs ^ResultSetMetaData rsmeta ^Integer i]
  (.getObject rs i))

An example column reader is provided -- clob-column-reader -- that still uses .getObject but will expand java.sql.Clob values into string (using the clob->string helper function):

    {:builder-fn (result-set/as-maps-adapter
                  result-set/as-maps
                  result-set/clob-column-reader)}

As of 1.1.569, the second style of adapter relies on with-column-value from RowBuilder (see below) and allows you to take complete control of the column reading process. This style takes a column-by-index-fn function, which is called with the builder itself, the ResultSet, and the column index, and is expected to read the raw column value from the result set and perform any and all processing on it, before returning it. The result is added directly to the current row with no further processing.

  • builder-adapter -- adapts any existing builder function with a new column reading function.

The default column-by-index-fn function behavior would be:

(defn default-column-by-index-fn
  [builder ^ResultSet rs ^Integer i]
  (result-set/read-column-by-index (.getObject rs i) (:rsmeta builder) i))

Because the builder itself is passed in, the vector of processed column names is available as (:cols builder) (in addition to the ResultSetMetaData as (:rsmeta builder)). This allows you to take different actions based on the metadata or the column name, as well as bypassing the read-column-by-index call if you wish.

The older as-*-adapter functions are now implemented in terms of this builder-adapter because with-column-value abstracts away how the new column's value is added to the row being built.

RowBuilder Protocol

This protocol defines five functions and is used whenever next.jdbc needs to materialize a row from a ResultSet as a Clojure data structure:

  • (->row builder) -- produces a new row (a (transient {}) by default),
  • (column-count builder) -- returns the number of columns in each row,
  • (with-column builder row i) -- given the row so far, fetches column i from the current row of the ResultSet, converts it to a Clojure value, and adds it to the row (for as-maps this is a call to .getObject, a call to read-column-by-index -- see the ReadableColumn protocol below, and a call to assoc!),
  • (with-column-value builder row col v) -- given the row so far, the column name, and the column value, add the column name/value to the row in the appropriate way: this is a low-level utility, intended to be used in builders (or adapters) that want to control more of the value handling process -- in general, with-column will be implemented by calling with-column-value,
  • (row! builder row) -- completes the row (a (persistent! row) call by default).

execute! and execute-one! call these functions for each row they need to build. plan may call these functions if the reducing function causes a row to be materialized.

ResultSet Protocol

This protocol defines three functions and is used whenever next.jdbc needs to materialize a result set (multiple rows) from a ResultSet as a Clojure data structure:

  • (->rs builder) -- produces a new result set (a (transient []) by default),
  • (with-row builder rs row) -- given the result set so far and a new row, returns the updated result set (a (conj! rs row) call by default),
  • (rs! builder rs) -- completes the result set (a (persistent! rs) call by default).

Only execute! expects this protocol to be implemented. execute-one! and plan do not call these functions.

Result Set Builder Functions

The as-* functions described above are all implemented in terms of these protocols. They are passed the ResultSet object and the options hash map (as passed into various next.jdbc functions). They return an implementation of the protocols that is then used to build rows and the result set. Note that the ResultSet passed in is mutable and is advanced from row to row by the SQL execution function, so each time ->row is called, the underlying ResultSet object points at each new row in turn. By contrast, ->rs (which is only called by execute!) is invoked before the ResultSet is advanced to the first row.

The result set builder implementation is also assumed to implement clojure.lang.ILookup such that the keys :cols and :rsmeta are supported and should map to the vector of column names that the builder will produce and the ResultSetMetaData object (which can be obtained from the ResultSet, if necessary). This is intended to allow plan and various builder adapters to access certain information that may be needed for processing results. The default builder implementations (for maps and arrays) are both records with fields rsmeta and cols (in addition to rs -- the ResultSet itself). The adapters provided in next.jdbc.result-set returned reified implementations that delegate field lookup to the underlying builder implementation.

The options hash map for any next.jdbc function can contain a :builder-fn key and the value is used as the row/result set builder function. The tests for next.jdbc.result-set include a record-based builder function as an example of how you can extend this to satisfy your needs.

Note: When next.jdbc cannot obtain a ResultSet object and returns {:next.jdbc/count N} instead, the builder function is not applied -- the :builder-fn option does not affect the shape of the result.

The options hash map passed to the builder function will contain a :next.jdbc/sql-params key, whose value is the SQL + parameters vector passed into the top-level next.jdbc functions (plan, execute!, and execute-one!).

There is also a convenience function, datafiable-result-set, that accepts a ResultSet object (and a connectable and an options hash map) and returns a fully realized result set, per the :builder-fn option (or as-maps if that option is omitted).

The array-based builders warrant special mention:

  • When used with execute!, the array-based builders will produce a data structure that is a vector of vectors, with the first element being a vector of column names and subsequent elements being vectors of column values in the same corresponding order. The order of column names and values follows the "natural" order from the SQL operation, as determined by the underlying ResultSet.
  • When used with execute-one!, the array-based builders will produce a single vector containing the column values in the "natural" SQL order but you will not get the corresponding column names back.
  • When used with plan, the array-based builders will cause each abstract row to represent a vector of column values rather than a hash map which limits the operations you can perform on the abstraction to just Associative (get with a numeric key), Counted (count), and Indexed (nth). All other operations will either realize a vector, as if by calling datafiable-row, or will fail if the operation does not make sense on a vector (as opposed to a hash map).

next.jdbc.optional

This namespace contains variants of the six as-maps-style builders above that omit keys from the row hash maps if the corresponding column is NULL. This is in keeping with Clojure's views of "optionality" -- that optional elements should simply be omitted -- and is provided as an "opt-in" style of rows and result sets.

ReadableColumn

As mentioned above, when with-column is called, the expectation is that the row builder will call .getObject on the current state of the ResultSet object with the column index and will then call read-column-by-index, passing the column value, the ResultSetMetaData, and the column index. That function is part of the ReadableColumn protocol that you can extend to handle conversion of arbitrary database-specific types to Clojure values. It is extensible via metadata so the value you return can have metadata specifying the implementation of read-column-by-index.

If you need more control over how values are read from the ResultSet object, you can use next.jdbc.result-set/as-maps-adapter (or next.jdbc.result-set/as-arrays-adapter, or the more low-level but more generic next.jdbc.result-set/builder-adapter) which takes an existing builder function and a column reading function and returns a new builder function that calls your column reading function (with the ResultSet object, the ResultSetMetaData object, and the column index -- or the builder itself, the ResultSet object, and the column index in the case of builder-adapter) instead of calling .getObject directly. Note that the as-* adapters still call read-column-by-index on the value your column reading function returns.

In addition, inside plan, as each value is looked up by name in the current state of the ResultSet object, the read-column-by-label function is called, again passing the column value and the column label (the name used in the SQL to identify that column). This function is also part of the ReadableColumn protocol.

The default implementation of this protocol is for these two functions to return nil as nil, a Boolean value as a canonical true or false value (unfortunately, JDBC drivers cannot be relied on to return unique values here!), and for all other objects to be returned as-is.

next.jdbc makes no assumptions beyond nil and Boolean, but common extensions here could include converting java.sql.Date to java.time.LocalDate and java.sql.Timestamp to java.time.Instant for example:

(extend-protocol rs/ReadableColumn
  java.sql.Date
  (read-column-by-label [^java.sql.Date v _]
    (.toLocalDate v))
  (read-column-by-index [^java.sql.Date v _2 _3]
    (.toLocalDate v))
  java.sql.Timestamp
  (read-column-by-label [^java.sql.Timestamp v _]
    (.toInstant v))
  (read-column-by-index [^java.sql.Timestamp v _2 _3]
    (.toInstant v)))

Remember that a protocol extension will apply to all code running in your application so with the above code all timestamp values coming from the database will be converted to java.time.Instant for all queries. If you want to control behavior across different calls, consider the adapters described above (as-maps-adapter, as-arrays-adapter, and builder-adapter, and think about using metadata to implement the rs/ReadableColumn protocol instead of extending it).

Note that the converse, converting Clojure values to database-specific types is handled by the SettableParameter protocol, discussed in the next section (Prepared Statements).