5. Data Stores

As explained in Section 4, a data store encapsulates a unit of logically related information. Many applications will store all of their related data in one data store (although some applications may use more than one data store). It is important to keep in mind that a query and rule can operate only on one data store; thus, all information that should be queried or reasoned within one unit should be loaded into one data store.

As explained in Section 4, a data store serves as a container for other kinds of objects:

  • tuple tables are data store components that store facts (see Section 6);

  • data sources can be registered with a data store to access external, non-RDF data (see Section 7);

  • OWL axioms and Datalog rules are used to specify rules of inference that are to be applied to the data loaded into the data store (see Section 10);

  • statistics modules summarize the data loaded into a data store in a way that helps query planning.

The behavior of a data store can be customized using various parameters and properties, as explained in Section 5.4.

5.1. Operations on Data Stores

The following list summarizes the operations on data stores available in the shell or via one of the available APIs.

  • A data store can be created on a server. To create a data store, one must specify the data store name and zero or more parameters expressed as key-value pairs.

  • A data store can be deleted on the server. RDFox® allows a data store to be deleted only if there are no active connections to the data store.

  • A data store can be saved to and subsequently loaded from a binary file. The file obtained in this way contains all data store content; thus, when a data store is loaded from a file, it is restored to exactly the same state as before saving. RDFox supports the following binary formats.

    • The standard format stores the data in a way that is more resilient to changes in RDFox implementation. This format should be used in most cases.

    • The raw format stores the data in exactly the same way as the data is stored in RAM. This format allows one to reconstruct the state of a data store exactly and is therefore useful when reporting bugs, but it is more likely to change between RDFox releases.

5.2. Organization of Data

The following sections explain how data is organized within a data store with details on how the data are stored, managed, imported, and queried.

5.2.1. Storage

By default, two in-memory tuple tables are created automatically when a data store is created. The first, DefaultTriples, has arity three and is designed to hold triples from an RDF dataset’s default graph. The second, Quads, has arity four and is designed to hold triples from all of the RDF dataset’s named graphs. If the default-graph-name (see Section 5.4.1.1) parameter is specified at data store creation time, the DefaultTriples table is not created.

5.2.2. Importation and DELETE/INSERT Clauses

By default, when performing import operations or evaluating SPARQL DELETE or INSERT clauses, RDFox maps default graph triples to entries in the DefaultTriples table and named graph triples (also known as quads) to entries in the Quads table. While named graph triples are always mapped to the Quads table, it is possible to override the default mapping for default graph triples so that they too are mapped to entries in the Quads table. The exact mechanisms for doing this are described next.

For import operations, the default mapping for default graph triples can be overridden by setting the target default graph parameter (see Section 8.2.1) of the operation. In this case, default graph triples are written to the Quads table with the specified graph name. When the target default graph parameter is not provided for the specific operation, but the data store’s default-graph-name parameter is set (see Section 5.4.1.1), the data store parameter is used as the graph name instead.

For SPARQL DELETE and INSERT clauses, overriding the default graph can be achieved by wrapping the triple patterns in the standard syntax for named graphs. When this syntax is not used, any setting for the default-graph-name parameter is used as the name of the graph that should receive all default triples.

When importing rules, default graph atoms are interpreted as references to the the DefaultTriples table unless the target default graph parameter is set for the import or the default-graph-name parameter is set for the data store. In the latter two cases, default graph atoms are interpreted as references to the named graph with the specified name.

When importing OWL, all axioms are applied to the DefaultTriples table unless the target default graph parameter is set for the import or the default-graph-name parameter is set for the data store. In the latter two cases, axioms are instead applied to the named graph with the specified name.

5.2.3. Mapping Triples and Quads to RDF Datasets for SPARQL WHERE/ASK

Before evaluating SPARQL WHERE clauses or ASK queries, it is necessary to determine the RDF dataset for the query. An RDF dataset comprises exactly one default graph plus zero or more named graphs. The default RDF dataset for an RDFox data store contains the triples stored in the data store’s DefaultTriples table in the default graph and all named graphs stored in the data store’s Quads table as its named graphs.

If the DefaultTriples table is not present when a query begins, the default RDF dataset instead contains the multiset of all triples from the data store’s Quads table in the default graph and, as above, all named graphs stored in the data store’s Quads table as its named graphs.

Regardless of whether the DefaultTriples table exists or not, the default behaviors described above can always be overridden using either FROM and FROM NAMED clauses in the query itself or, in the case of the SPARQL protocol, the default-graph-uri and named-graph-uri request parameters.

5.2.4. Compaction and Dead Fact Removal

Facts added to any in-memory tuple table can be deleted using any of the APIs provided by RDFox, and all facts can be removed from a data store using a dedicated clear operation. Facts deleted in any of these two ways will not be taken into account during reasoning or query evaluation. However, it is important to bear in mind that fact deletion typically only marks the fact as deleted and does not reclaim any resources used to store the fact. Facts that have been marked for deletion are sometimes called dead, and all other facts are called live. If a fact is added to a data store that already contains the same fact marked for deletion, the fact is resurrected – that is, the deletion flag is removed and the fact becomes live. As a consequence of these implementation choices, an application that keeps adding and deleting batches of distinct facts will eventually run out of memory, even if the number of facts that are live at any point in time is bounded.

Memory occupied by dead facts can be reclaimed by compacting a data store. Roughly speaking, this operation removes all dead facts from memory, rebuilds all indexes used for query evaluation, and updates all statistics used during query planing. This operation increments the version of the data store even though the data store content does not change conceptually. The memory used by a data store after compaction should be roughly the same as if the data was imported into a fresh data store. It is thus generally good practice to compact a data store after periods of substantial data deletion.

If a data store is persistent (see Section 13), compacting it also saves a fresh snapshot of its entire content to disk and removes older snapshots and deltas to free up disk space. On startup, RDFox loads data stores by first loading their most recent snapshot and then replaying the changes in all subsequent deltas in order. Compacting regularly helps to speed up server restart by minimizing the number of deltas that need to be replayed after the most recent snapshot is loaded.

Since compaction completely reorganizes the data in a data store, it can be performed only in exclusive mode – that is, if no readers and no writers are active on the data store. Moreover, compaction requires reindexing the data store, which can take considerable time if the data store contains a lot of data. Thus, in applications where the data changes frequently, it is prudent to schedule compaction at times of low use.

For persistent data stores, compaction will also temporarily increase the disk usage of the volume containing the server directory by the size of the data store’s new snapshot, which must be safely written to disk before older snapshots and deltas can be removed. Because of this, operators should ensure that sufficient disk space is available before initiating compaction.

Because compaction can be a heavy-weight operation, particularly when a new snapshot of the data store is saved to disk, RDFox also supports dead fact removal, which can be seen as a lighter form of compaction. Unlike compaction, it is applied only to tuple tables that contain more than 50% dead facts, and it does not save a new snapshot of the data store. Thus, while dead fact removal also requires exclusive access to the data store, the exclusive lock will typically be held for shorter periods of time and so this operation is less likely to disturb client operation, and there is no associated increase in disk usage.

An operator can explicitly initiate data store compaction using any of the available APIs. Moreover, RDFox can be configured to perform compaction and dead fact removal automatically as follows.

  • The auto-compact-after data store property (see Section 5.4.2) can be used to configure a data store to be automatically compacted after a fixed number of committed transactions. For example, if the auto-compact-after data store property is set to 5, then the data store will be compacted automatically if (a) the last compaction occurred at least five transactions ago and (b) no readers and no writers are active on the data store when the transaction is committed. If a reader or writer is active, compaction will be attempted when the next transaction is committed. This option should be used only when compaction overhead is unlikely to adversely affect the query and update rate of an RDFox instance. The default value for the auto-compact-after data store property is never – that is, automatic compaction is switched off by default.

  • If the value of the remove-dead-facts data store property (see Section 5.4.2) is set to auto, then dead fact removal will be performed every time when a transaction is committed or rolled back. Specifically, provided that no readers and no writers are active on a data store at the point of commit or rollback, dead facts will be removed from every tuple table that contains more than 50% of dead facts. If an exclusive lock cannot be obtained, dead fact removal will be attempted again on the next commit or rollback. Dead fact removal is switched on by default as this should pose no problems for many applications. However, the need to acquire an exclusive lock can introduce an occasional slowdown in query response times, in which case the remove-dead-facts data store property should be set to off.

5.3. Supported Data Types for Literals

As well as IRIs and blank nodes, RDFox data stores can store literal values in the following formats:

Datatype

http://www.w3.org/2000/01/rdf-schema#Literal

http://www.w3.org/2001/XMLSchema#anyURI

http://www.w3.org/2001/XMLSchema#string

http://www.w3.org/1999/02/22-rdf-syntax-ns#PlainLiteral

http://www.w3.org/2001/XMLSchema#boolean

http://www.w3.org/2001/XMLSchema#dateTime

http://www.w3.org/2001/XMLSchema#dateTimeStamp

http://www.w3.org/2001/XMLSchema#time

http://www.w3.org/2001/XMLSchema#date

http://www.w3.org/2001/XMLSchema#gYearMonth

http://www.w3.org/2001/XMLSchema#gYear

http://www.w3.org/2001/XMLSchema#gMonthDay

http://www.w3.org/2001/XMLSchema#gDay

http://www.w3.org/2001/XMLSchema#gMonth

http://www.w3.org/2001/XMLSchema#duration

http://www.w3.org/2001/XMLSchema#yearMonthDuration

http://www.w3.org/2001/XMLSchema#dayTimeDuration

http://www.w3.org/2001/XMLSchema#double

http://www.w3.org/2001/XMLSchema#float

http://www.w3.org/2001/XMLSchema#decimal

http://www.w3.org/2001/XMLSchema#integer

http://www.w3.org/2001/XMLSchema#nonNegativeInteger

http://www.w3.org/2001/XMLSchema#nonPositiveInteger

http://www.w3.org/2001/XMLSchema#negativeInteger

http://www.w3.org/2001/XMLSchema#positiveInteger

http://www.w3.org/2001/XMLSchema#long

http://www.w3.org/2001/XMLSchema#int

http://www.w3.org/2001/XMLSchema#short

http://www.w3.org/2001/XMLSchema#byte

http://www.w3.org/2001/XMLSchema#unsignedLong

http://www.w3.org/2001/XMLSchema#unsignedInt

http://www.w3.org/2001/XMLSchema#unsignedShort

http://www.w3.org/2001/XMLSchema#unsignedByte

5.4. Data Store Configuration

The behavior of a data store can be customized using options expressed as key-value pairs. These options are arranged into the following two groups.

  • Data store parameters are options that are set when a data store is created and cannot be changed subsequently. They are supplied to various APIs that allow data store creation, and they generally govern issues such as data store capacity or the persistence scheme used. Data store parameters are described in more detail in Section 5.4.1.

  • Data store properties are options that can change throughout the lifetime of a data store. RDFox provides various APIs for retrieving and changing data store properties; moreover, if a data store is persisted, any changes to data store properties are persisted too. Data store properties are described in more detail in Section 5.4.2.

5.4.1. Data Store Parameters

Data store parameters are key-value pairs that are determined when a data store is created and that cannot be changed subsequently.

5.4.1.1. default-graph-name

The default-graph-name option can be set to any IRI to activate RDFox’s alternative handling of default graph data. See Section 5.2 for a full description. By default, this parameter is unset.

5.4.1.2. equality

The equality option determines how RDFox deals with the semantics of equality, which is encoded using the owl:sameAs property. This option has the following values.

  • off: There is no special handling of equality and the owl:sameAs property is treated as just another property. This is the default if the equality option is not specified.

  • noUNA: The owl:sameAs property is treated as equality, and the Unique Name Assumption is not used — that is, deriving an equality between two IRIs does not result in a contradiction. This is the treatment of equality in OWL 2 DL.

  • UNA: : The owl:sameAs property is treated as equality, but interpreted under UNA — that is, deriving an equality between two IRIs results in a contradiction, and only equalities between an IRI and a blank node, or between two blank nodes are allowed. Thus, if a triple of the form <IRI₁, owl:sameAs, IRI₂> is derived, RDFox detects a clash and derives <IRI₁, rdf:type, owl:Nothing> and <IRI₂, rdf:type, owl:Nothing>.

In all equality modes (i.e., all modes other than off), distinct RDF literals (e.g., strings, numbers, dates) are assumed to refer to distinct objects, and so deriving an equality between the distinct literals results in a contradiction.

Note RDFox will reject rules that use negation-as-failure or aggregation in all equality modes other than off.

5.4.1.3. init-resource-capacity

The value of the init-resource-capacity option is an integer that is used as a hint to the data store specifying the number of resources that the store will contain. This hint is used to initialize certain data structures to the sizes that ensure faster importation of data. The actual number of resources that a data store can contain is not limited by this option: RDFox will resize the data structures as needed if this hint is exceeded.

5.4.1.4. init-tuple-capacity

The value of the init-tuple-capacity option is an integer that is used as a hint to the data store specifying the number of tuples that the store will contain. This hint is used to initialize certain data structures to the sizes that ensure faster importation of data. The actual number of tuples that a data store can contain is not limited by this option: RDFox will resize the data structures as needed if this hint is exceeded.

5.4.1.5. max-data-pool-size

The value of the max-data-pool-size option is an integer that determines the maximum number of bytes that RDFox can use to store resource values (e.g., IRIs and strings). Specifying this option can reduce significantly the amount of virtual memory that RDFox uses per data store.

5.4.1.6. max-resource-capacity

The value of the max-resource-capacity option is an integer that determines the maximum number of resources that can be stored in the data store. Specifying this option can reduce significantly the amount of virtual memory that RDFox uses per data store.

5.4.1.7. max-tuple-capacity

The value of the max-tuple-capacity option is an integer that determines the maximum number of tuples that can be stored by the in-memory tuple tables of a data store. Specifying this option can reduce significantly the amount of virtual memory that RDFox uses per data store.

5.4.1.8. persistence

The persistence option controls how RDFox persists data contained in a data store. The option can be set to off to disable persistence for individual data stores contained within a persistent server. The option can also be set to the same value as the persistence option of the server that contains the data store but, as this is the default value, this is not necessary.

5.4.1.9. quad-table-type

The parameter quad-table-type, determines the type of the table Quads, which is used to store named graph triples. The available options are quad-table-lg, the default value, and quad-table-sg. Tuple table type quad-table-lg uses indexing suitable for the typical use cases where named graphs contain non-trivial amount of facts. In contrast, the type quad-table-sg uses indexing suitable for the rare cases where each graph consists of a very few triples.

5.4.1.10. swrl-negation-as-failure

The swrl-negation-as-failure option determines how RDFox treats ObjectComplementOf class expressions in SWRL rules.

  • off. SWRL rules are interpreted under the open-world assumption and SWRL rules featuring ObjectComplementOf are rejected. This is the default value.

  • on. SWRL rules are interpreted under the closed-world assumption, as described in Section 10.7.3.

5.4.1.11. type

The type option determines the storage scheme used by the data store. The value determines the maximum capacity of a data store (i.e., the maximum number of resources and/or facts), its memory footprint, and the speed with which it can answer certain types of queries. The following data store types are currently supported:

  • parallel-nn (default)

  • parallel-nw

  • parallel-ww

In suffixes nn, nw, and ww, the first character determines whether the system uses 32-bit (n for narrow) or 64-bit (w for wide) unsigned integers for representing resource IDs, and the second character determines whether the system uses 32-bit (n) or 64-bit (w) unsigned integers for representing triple or quad IDs. Thus, an nw store can contain at most 4 × 109 resources and at most 1.8 × 1019 triples.

5.4.2. Data Store Properties

Data store parameters are key-value pairs that can be modified at any point during a data store’s lifetime using any of the provided RDFox APIs (see Section 16) or the dsprop shell command (see Section 15.2.15).

5.4.2.1. auto-compact-after

The auto-compact-after data store property determines whether and how frequently a data store is compacted automatically. The allowed values are as follows.

  • never: The data store is never compacted automatically. (However, compaction can be requested explicitly using any of the provided APIs.). This is the default.

  • an integer n: The data store is compacted after n transactions are successfully committed, provided that the data store has no active readers or writers at the point of commit.

5.4.2.2. auto-update-statistics-mode

The auto-update-statistics-mode data store property governs how RDFox manages statistics about the data loaded into the system. RDFox uses these statistics during query planning in order to identify an efficient plan, so query performance may be suboptimal if the statistics are not up to date. The allowed values are as follows.

  • off: Statistics are never updated automatically, but they can be updated manually using the stats update command or via one of the available APIs.

  • balanced: The cost of updating the statistics is balanced against the possibility of using outdated statistics. This is the default.

  • eager: Statistics are updated after each operation that has the potential to invalidate the statistics (e.g., importing data).

5.4.2.3. base-iri

The base-iri data store property contains an IRI that is used as the base when importing data or evaluating queries with no explicitly set base IRI. All relative IRIs in such cases are resolved against the value of this property in order to obtain an absolute IRI. The defalt value is https://rdfox.com/default-base-iri/.

5.4.2.4. errors-in-bind

The errors-in-bind data store property governs how errors encountered during the evaluation of BIND expressions are handled in queries. As an example, consider SPARQL query SELECT ?X ?Y WHERE { BIND(2 + "a" AS ?X) . ?Y :R ?X }; here, the expression 2 + "a" raises an error because strings cannot be added to integers.

  • If the value of errors-in-bind is equal to standard-compliant, then variable ?X is unbound after evaluating the BIND expression, and the query becomes equivalent to SELECT ?X ?Y WHERE { ?Y :R ?X }. This is the default behavior.

  • If the value of errors-in-bind is equal to skip, then the attempt to bind ?X to an error is skipped and the query does not return any answers.

The default behavior is compliant with the SPARQL 1.1 standard, but it can be counterintuitive by silently accepting errors, and it can also lead to inefficient query plans. The value of this data store property can be overridden during query evaluation by passing the query parameter with the same name.

5.4.2.5. invalid-literal-policy

The invalid-literal-policy data store property governs how RDFox handles invalid literals in imported data, queries, and updates.

  • error: Invalid literals in the input are treated as errors, and so input containing such literals cannot be processed. This is the default.

  • as-string: During import, invalid literals are converted to string literals and a warning is emitted alerting the user to the fact that the value was converted. In queries and updates, invalid literals are converted to strings, but no warning is emitted.

  • as-string-silent: Invalid literals are converted to string literals in the import data, queries, and updates, but without emitting a warning.

5.4.2.6. max-backward-chaining-depth

The max-backward-chaining-depth data store property determines the recursion depth of the backward chaining phase in the incremental reasoning that RDFox uses to update the materialization. The value can be a nonnegative integer, or unbounded (the default). The value of this property should be changed only if so directed by Oxford Semantic Technologies support personnel.

5.4.2.7. max-threads-used

The max-threads-used data store property determines how many threads of the server’s thread pool this data store is allowed to use for tasks such as importation or reasoning. The value can be a nonnegative integer, or all-available; the latter is the default and it states that a data store can use as many threads of the server’s thread pool as are currently available.

5.4.2.8. property-paths-cardinality

The property-paths-cardinality data store property governs the evaluation of property paths in queries. As an example, consider SPARQL query SELECT ?X ?Y WHERE { ?X :R/:S ?Y }.

  • If the value of property-paths-cardinality is equal to standard-compliant, then the query is equivalent to SELECT ?X ?Y WHERE { ?X :R ?Z . ?Z :S ?Y } — that is, the query can return the same pair of values for ?X and ?Y more than once. This is the default.

  • If the value of property-paths-cardinality is equal to distinct, then the query is equivalent to SELECT DISTINCT ?X ?Y WHERE { ?X :R ?Z . ?Z :S ?Y } — that is, the query can return a pair of values for ?X and ?Y at most once.

Elimination of duplicate values is required by the standard whenever a property path contains *, +, or ?. Thus, the distinct value for the property-paths-cardinality data store property presents a more uniform and consistent semantics for SPARQL 1.1 property paths; moreover, complex property paths can often be evaluated more efficiently with the distinct option. The value of this data store property can be overridden during query evaluation by passing the query parameter with the same name.

5.4.2.9. query-planning-algorithms

The query-planning-algorithms data store property governs which algorithms are used to plan a query. The value of this property should be changed only if so directed by Oxford Semantic Technologies support personnel. The value of this data store property can be overridden during query evaluation by passing the query parameter with the same name.

5.4.2.10. remove-dead-facts

The remove-dead-facts data store property specifies whether dead fact removal is applied automatically after a transaction is committed or rolled back. The allowed values are auto (the default), and off.

5.4.2.11. query-validation

The query-validation data store property specifies whether RDFox should reject queries that conform to the SPARQL 1.1 standard in a strict sense, but exhibit a pattern that was often found to represent a user error. The value standard-compliant instructs RDFox to closely follow the SPARQL 1.1 specification, whereas the value strict instructs RDFox to perform further checks, some of which are listed below. The default value is strict. The value of this data store property can be overridden during query evaluation by passing the query parameter with the same name.

The following list presents some of the types of queries that RDFox will reject in strict mode, all of which represent common mistakes.

  • Query SELECT * WHERE { BIND(?X + 2 AS ?Y) . ?Z :R ?X } will be rejected because, as per the SPARQL 1.1 standard, BIND(?X + 2 AS ?Y) must be evaluated before triple pattern ?Z :R ?X, and so variable ?X is necessarily unbound when BIND is evaluated. Such a query should usually be written as SELECT * WHERE { ?Z :R ?X . BIND(?X + 2 AS ?Y) }.

  • Query SELECT * WHERE { ?X :R ?Y . FILTER(?Z > 2) } will be rejected because variable ?Z occurs in FILTER(?Z > 2), but not in ?X :R ?Y. Variable ?Z is thus necessarily unbound. Such queries usually contain a typo: this query was likely meant to be SELECT * WHERE { ?X :R ?Y . FILTER(?X > 2) }.

  • Query SELECT * WHERE { ?X :R ?Y . OPTIONAL { ?X :S ?Y . FILTER(?Z > 2) } will be rejected because variable ?Z in FILTER(?Z > 2) occurs in neither ?X :R ?Y nor ?X :S ?Y, and so the variable is necessarily unbound. Such queries usually contain a typo just as in the previous item.

  • Query SELECT * WHERE { GRAPH :G { ?S ?P ?O } } will be rejected if either the dataset does not include the named graph :G, or the user does not have permissions to read the named graph :G. SPARQL 1.1 requires such queries to be evaluated as if named graph :G were empty; however, this does not support detecting errors (e.g., if :G was mistyped in the query).

  • Query SELECT * WHERE { ?S ?P ?O } ORDER BY ?Z will be rejected because variable ?Z is used in ORDER BY, but is not bound in the main part of the query.

  • Query SELECT ?X ?Y WHERE { ?X :R ?Y } GROUP BY ?X will be rejected because the query uses GROUP BY, but variable ?Y occurs neither in GROUP BY nor in an aggregation function. SPARQL 1.1 standard stipulates that the value of ?Y should be sampled in such cases — that is the query should be evaluated as if it was written as SELECT ?X (SAMPLE(?Y1) AS ?Y) WHERE { ?X :R ?Y1 } GROUP BY ?X. However, such implicit, significant rewriting of the query often tends to hide a typo or an omission of ?Y from GROUP BY.

  • Query SELECT ?X ?Z WHERE { ?X :R ?Y } will be rejected because variable ?Z is selected, but not bound in the query. Variable ?Z is thus necessarily unbound, which usually indicates a typo.

5.4.2.12. user-blank-node-import-policy

If the user-blank-node-import-policy data store property is set to rename-apart, then user-defined blank nodes imported from distinct files are renamed apart during the importation process; hence, importing data merges blank nodes according to the RDF specification. There is no way to control the process of renaming blank nodes, which can be problematic in some applications. Because of that, the default value of this option is keep-unchanged, which ensures that the data is imported ‘as is’. Regardless of the state of this option, autogenerated blank nodes (i.e., blank nodes obtained by expanding [] or (...) in Turtle files) are always renamed apart.