Labeling versioned hierarchical data

CROSS REFERENCE TO RELATED APPLICATIONS

0001This application is related to, and claims benefit to and priority of, U.S. Patent Application Ser. No. 61/725,783, filed on Nov. 13, 2012, the contents of which are hereby incorporated by reference in their entirety for all purposes.

BACKGROUND

0002Many business operations feature hierarchical data. For example, business users need to manage human resource (HR) hierarchies, which model the relationship between employees. Most relational database systems do not provide efficient support for hierarchical data. Efficient query support of database hierarchies and of prior database states requires new techniques for maintaining and versioning such hierarchical data.

BRIEF DESCRIPTION OF THE DRAWINGS

0003<figref idref="DRAWINGS">FIG. 1</figref> is a block diagram of a computing system according to some embodiments.

0004<figref idref="DRAWINGS">FIG. 2</figref> illustrates a human resources hierarchy and its interval encoding according to some embodiments.

0005<figref idref="DRAWINGS">FIG. 3</figref> illustrates changes to a hierarchy and resulting changes to its interval encoding according to some embodiments.

0006<figref idref="DRAWINGS">FIG. 4</figref> illustrates an interval delta according to some embodiments.

0007<figref idref="DRAWINGS">FIG. 5</figref> illustrates a source version tree and a target version tree according to some embodiments.

0008<figref idref="DRAWINGS">FIG. 6</figref> illustrates determination of an interval delta based on two interval encodings according to some embodiments.

0009<figref idref="DRAWINGS">FIG. 7</figref> determination of an interval delta based on hierarchy updates according to some embodiments according to some embodiments.

0010<figref idref="DRAWINGS">FIG. 8</figref> illustrates an accumulation tree according to some embodiments.

0011<figref idref="DRAWINGS">FIG. 9</figref> illustrates rules for updating node values after a left node rotation according to some embodiments.

0012<figref idref="DRAWINGS">FIG. 10</figref> illustrates swapping two translation ranges according to some embodiments.

0013<figref idref="DRAWINGS">FIG. 11</figref> illustrates the size of a first ten interval deltas of a version history according to some embodiments.

0014<figref idref="DRAWINGS">FIG. 12</figref> illustrates merging of two interval deltas according to some embodiments.

0015<figref idref="DRAWINGS">FIG. 13</figref> is a block diagram of a computing system according to some embodiments.

DESCRIPTION

0016The following description is provided to enable any person in the art to make and use the described embodiments. Various modifications, however, will remain readily apparent to those in the art.

0017<figref idref="DRAWINGS">FIG. 1</figref> is a block diagram of database <b>100</b> according to some embodiments. <figref idref="DRAWINGS">FIG. 1</figref> represents a logical architecture for describing systems according to some embodiments, and actual implementations may include more or different components arranged in other manners.

0018Database <b>100</b> may comprise any query-responsive data source or sources that are or become known, including but not limited to a structured-query language (SQL) relational database management system. Database <b>100</b> may comprise a relational database, a multi-dimensional database, an eXtendable Markup Language (XML) document, or any other data storage system storing structured and/or unstructured data. The data of database <b>100</b> may be distributed among several relational databases, multi-dimensional databases, and/or other data sources. Embodiments are not limited to any number or types of data sources. For example, database <b>100</b> may comprise one or more OnLine Analytical Processing (OLAP) databases (i.e., cubes), spreadsheets, text documents, presentations, etc.

0019Database <b>100</b> may implement an “in-memory” database, in which volatile (e.g., non-disk-based) storage (e.g., Random Access Memory) is used both for cache memory and for storing the full database during operation, and persistent storage (e.g., one or more fixed disks) is used for offline persistency and maintenance of database snapshots. Alternatively, volatile storage may be used as cache memory for storing recently-used data, while persistent storage stores the full database. In some embodiments, the data of database <b>100</b> comprises one or more of conventional tabular data, row-based data stored in row format, column-based data stored in columnar format, and object-based data. Database <b>100</b> may also or alternatively support multi-tenancy by providing multiple logical database systems which are programmatically isolated from one another. Moreover, the data of database <b>100</b> may be indexed and/or selectively replicated in an index to allow fast searching and retrieval thereof.

0020Database <b>100</b> may store metadata describing regarding the structure, relationships and meaning of the data stored within database <b>100</b>. This information may include data defining the schema of database tables stored within database <b>100</b>. A database table schema may specify the name of the database table, columns of the database table, the data type associated with each column, and other information associated with the database table.

0021Database <b>100</b> includes, among other data, versioned hierarchical data <b>110</b>. Conserving database versions allows the execution of queries not only on the latest state of the database but also on older states. If the HR hierarchy of a large company is to be versioned on a daily basis for several years, the result is a versioned hierarchy with millions of nodes and thousands of versions. Other applications may require even finer-grained version control, potentially resulting in millions of versions.

0022Database engine <b>112</b> performs administrative and management functions for database <b>100</b>. Such functions may include snapshot and backup management, indexing, optimization, garbage collection, and/or any other database functions that are or become known. For example, database engine <b>112</b> may operate as described herein to generate interval encodings <b>114</b> and interval deltas <b>116</b>.

0023Interval encodings <b>114</b> are used for indexing hierarchical data <b>110</b>. For example, each node of a hierarchy is assigned a constant number of labels and certain queries can be answered by only considering the node labels. A prominent known labeling scheme is the nested intervals (NI) labeling scheme, in which each node is labeled with an interval [low, high] and the interval for a node v is a proper subinterval of interval of the parent node of v. <figref idref="DRAWINGS">FIG. 2</figref> illustrates an example HR hierarchy and its corresponding interval encoding.

0024The below-described DeltaNI index applies a nested intervals (NI) scheme to versioned hierarchies. Some embodiments of the index are efficient in space and time by representing only a base version as a fully materialized NI encoding; other versions are represented by deltas (e.g., interval deltas <b>116</b>) which transform the interval bounds between different versions. By grouping the deltas in an exponential fashion, some embodiments allow executing queries in each version of a history of n versions while applying at most log n deltas. Many update operations, including but not limited to moving or deleting whole ranges of nodes, can be reduced to a simple swap operation in the NI encoding. As many query processing techniques that work on top of NI labeling (as, e.g., the staircase join) have already been proposed, the DeltaNI index can be used as a building block for processing various kinds of queries.

0025For purposes of the present description, a hierarchy is defined as an ordered, rooted forest T=(V,E), which is a directed graph consisting of a set of nodes (vertices) V and an edge relation E<u style="single">⊂</u>V×V. The graph is free of circles and the edge relation is injective. d denotes the maximum depth of a hierarchy, i.e., the longest path from any node to its root, and n=|V| denotes the number of nodes in a tree. Applications can also limit a hierarchy to be a tree instead of a forest (i.e., only one root exists) or may specify that the order among siblings is not important and thus is not made visible to the user. Each vertex may contain a fixed number of attributes while edges may not carry attributes in some embodiments.

0026A straightforward representation for hierarchies in relational databases is to save the key of the parent node of each node in a parent column. The tree root has NULL as parent. Such encoding may yield rather bad performance for queries that rely on the recursive structure of a hierarchies, like whether node X is an ancestor of node Y. This query would require a backward search from Y which is naïvely implemented as a recursive self-join resulting in n self-joins over the parent column, where n is the length of the path from Y to the root.

0027In order to speed up queries which work on recursive properties of the hierarchy the parent column is dropped and instead the hierarchy is represented by a relation, in which each row represents one node storing its attributes, and an index (i.e., the DeltaNI index described herein) which encodes the structure of the hierarchy.

0028As previously mentioned, the index may be built upon the NI labeling scheme (also referred to as interval encoding, NI encoding, or range-based encoding). Here, each node is represented by the integer interval [low, high] (also referred to as [left, right] or [begin,end]). The encoding can be obtained by a depth-first traversal. Whenever a node is visited first, before its children are visited (pre-order) its low bound is assigned from a counter variable which starts at zero and is incremented whenever a bound is assigned. After the low bound has been assigned and the child nodes are visited, the high bound is assigned from the counter variable (post-order). If a node v₂is a descendant of another node v₁its [low,high] interval is a proper sub-interval of v₁'s interval. Formally, an ancestor query “is v₁ancestor of v₂” can be answered positively, iff n₁.low<n₂.low and n₁.high>n₂.high. Similarly, a preceding query “is n₁a preceding node of n₂” is answered positively, iff the interval of n₁ends before the interval of n₂starts, i.e., n₁.high<n₂.low. Because these queries require at most two integer comparisons, their time complexity is ο(1), which may result in a large performance gain compared to repetitive self-joining.

0029Hierarchy Queries and Updates

0030An index according to some embodiments may yield a fully-featured NI encoding for each version of the hierarchy. Consequently, all kinds of queries which a usual NI encoding can answer for a non-versioned hierarchy can be efficiently also answered for versioned hierarchies using a DeltaNI index described herein. For example, the DeltaNI index can be used as a basis for the staircase join, the Stack-Tree join, or the TwigStack join in order to answer XPath-style queries.

0031<figref idref="DRAWINGS">FIG. 2</figref> illustrates a human resources hierarchy to provide an example of a query which can be executed directly on an DeltaNI index. The query requests the size of the subtree rooted at a node v, which is executed by calculating (v.high−v.low−1)/2. With respect to <figref idref="DRAWINGS">FIG. 2</figref>, a subtree query may be: “How many employees are (transitively) supervised by Adam (node A)”. Using the interval encoding, the answer is (15−0−1)/2=7. The index may also be used by a database system to answer a wide range of other complex queries efficiently. In the versioned case, the query would be extended to work on a certain version, e.g., “How many employees are (transitively) supervised by Adam in Version V_x”.

0032Updating a hierarchy consists of adding, removing, or moving nodes in the hierarchy. Altering the attributes of the nodes is not discussed below since the attributes are stored in usual relations and thus need no extra handling. The following update operations may be supported in some embodiments:

0033insertBefore(b): Inserts a new node before bound b.

0034moveSiblingRangeBefore(v,v′,b): v is a left sibling of v′ or v=v′. The operations moves all siblings between v and v′ (inclusively) and their descendants before bound b.

0035deleteSiblingRange(v,v′): v is a left sibling of v′ or v=v′. The operations deletes all siblings between v and v′ (inclusively) and their descendants.

0036The above-defined set of update operations is powerful, as the operations allow not only single node insertion and deletion but also subtree deletion and the moving of nodes, subtrees, and even whole ranges of siblings. These operations may be important in many use cases: For example, a division in a human resources hierarchy receiving a new head can be modeled by simply moving all nodes in that division below the new head with a single operation. With insert and delete operations only, this would result in one delete and one insert per person in this division. As most hierarchies are rather flat but have a large fan-out, this would result in a high update cost and the resulting delta would contain many operations also yielding increased space consumption. Consequently, the above update operations are useful in a wide applicability of a hierarchy index. An index according to some embodiments supports all these updates in worst case logarithmic time.

0037Versioned Hierarchies

0038A version history V₀, V₁, . . . , V_nof a hierarchy depicts certain states of that hierarchy and allows queries in each version. Updates are allowed in the latest version. A new branch of version histories can be created based on any existing version. With the existence of multiple branches, updates can be performed on the latest version of each branch. Although branching is supported by the index, a linear version history will be assumed herein for brevity. Supporting branching is straightforward without any adjustments to the index being necessary. Some applications might create a new version with each new change while others might create new versions on a regular basis (e.g., every day).

0039Although versioning of the hierarchy structure is considered herein, other versioning considerations, like the versioning of changing node attributes, can be achieved by traditional techniques for versioning (or transaction-time) relational databases, as these attributes are stored in a usual relational table.

0040Use Cases for DeltaNI

0041One use case for the index according to some embodiments is the version control of hierarchical data. Another possible use case includes transaction-time temporal hierarchies. The index (as any other version control approach) can be directly used for this purpose. An additional lookup data structure (for example, a search tree) which maps time intervals to versions is maintained, allowing to find the version that corresponds to a certain timestamp. The maintenance of such a data structure is straightforward and is therefore not covered here. Instead of working with XML, we assume general hierarchies that subsume XML. Thus, the index can also be used for the management of versioned XML data.

0042The NI encoding is by default not dynamic (i.e., not efficiently updatable), since an update needs to update O(n) bounds on average. In contrast, the DeltaNI index may be used as an efficiently-updatable NI labeling scheme for non-versioned databases. Since a delta can be updated efficiently, using a single delta for representing all incoming updates is sufficient for making an NI encoding dynamic. Finally, the deltas in this approach can also be used for logging purposes, as a delta accurately describes a set of changes.

0043Efficiently storing the version history of a hierarchy may consist of saving one or more base versions explicitly using interval encoding and saving all other versions as interval deltas only. This may allow for space-efficient compression of the version history while still allowing efficient queries.

0044An interval delta δ: N→N is defined as a function mapping interval bounds from a source version V to a target version V′. When needed, the source and target versions of a delta are explicitly specified using the notation δ_V→V′. Given an interval bound b of a node in V, δ_V→V′(b) yields the corresponding bound in V′. The inverse function δ_V→V′⁻¹=δ_V→V′ can be used for mapping back from bounds in the target version to bounds in the source version. We denote the interval encoding of the source version as source space and the one of the target version as target space. Thus, δ is a function mapping from the source to the target space.

0045The full interval encoding of the target version can be obtained by applying δ to all interval bounds of the source version. However, the delta can also be used to answer certain queries without computing the target version intervals completely. For example, consider two nodes v and v′ with known interval bounds in the source version V. Interval encoding allows a test of whether v′ is a descendant of v in version V by applying the condition v.low<v′.loŵv.high>v′.high. Using the interval delta δ from V to V′, the same query can be answered in the target version by applying the condition δ(v.low)<δ(v′.low)̂A δ(v.high)>δ(v′.high). Thus, if we have a base version V with known interval bounds and a version delta δ_V→V′, we can answer queries in V′ as well. To save a version history thoroughly, it suffices to save one base version (e.g., the first version) and then save only deltas from the base version to each other version.

0046When using interval deltas to represent the version history of a hierarchy, not all nodes may have existed in the base version V. Instead, they may have been added in a later version. These nodes do not have any bounds in the base version, thus computing their bounds in other versions V′ using δ_V→V′, is impossible. In addition, there might be nodes which were deleted in intermediate versions. It is unclear what the result of δ_V→V′(b) should be if b belongs to a node which was deleted in a version between V and V′. To handle insertions and deletions consistently, we make the following enhancements, which we call active region approach: For each version V of the history, the maximum bound value in that version, denoted as max(V), is stored. By definition, any bound value greater than max(V) does not exist in version V (i.e., “is not active”). In addition, for every base version V, we define |V| as the number of bounds stored in V also including bounds which do not exist in that version, that is, bounds which are greater than max(V). These enhancements allow us to model bounds that do not exist in a version. Consider a base version V and a version V′ which adds a new node v with bounds [v.low, v.high]. This node insertion is modeled by adding the two bounds b₁=|V| and b₂=|V|+1 into the base version V (which also increments |V| by two) but without increasing max(V), because b₁and b₂do not exist in V. To yield the correct result in V′, the delta is defined correspondingly: δ_V→V′(b₁)=v.low and δ_V→V′(b₂)=v.high. Then, max(V′) is incremented by two because this version now contains two more bounds. A node deletion in a version V′ can simply be achieved by moving the bounds of the node to be deleted past max(V′). We define the interval (max(V), ∞) as the inactive region and the interval [0, max(V)] as the active region of version V. Bounds in the active region represent bounds that exist in V and bounds in the inactive region represent bounds of nodes that do not exist in V. We call bounds in the active region active bounds and bounds in the inactive region virtual bounds, because they do not exist, actually. When using the active region approach, the test whether a node v exists in a version V to which a delta δ exists can be implemented by checking whether the lower bound of v (and thus also the upper bound) is active, i.e., δ(v.low)≦max(V).

0047Each node is uniquely identified by its bounds in the base version, since these bounds will never be updated. Thus, the bounds pose a durable numbering for the nodes in a versioned hierarchy. Given a bound b in a version V′, one can obtain the node to which b belongs by applying reverse deltas to b transforming the bound back to the base version and looking up the corresponding node there.

0048Interval deltas, together with the active region approach, can be used to model the version history of a hierarchy. We present an efficient representation of the interval delta function δ together with algorithms for building such delta from two explicitly represented versions or from a base version and a sequence of update operations on that version. We also introduce a technique for reducing the size of all deltas which need to be stored to represent a large version history with thousands or even millions of versions compactly.

0049To render the interval delta approach feasible, the resulting delta representation should be efficient in space and time. A reasonable space complexity requirement for a delta δ is ο(c), where c is the number of atomic changes like node insertion, node deletion, or subtree relocation which led from the source to the target version. In the worst case, this is also the best possible bound, because each change must be represented somehow requiring at least a constant amount of space. A reasonable upper bound for the time complexity of δ, i.e., the time it takes to compute δ(b) for any interval bound b in the source version, is ο(log(c)). Any higher non-logarithmic bound would make the approach infeasible for deltas containing a large amount of changes. Our approach satisfies both mentioned complexity bounds. The space and time complexity of our δ representation grows only with respect to the number of changes between the source and the target version. Especially, the complexities do not grow with the number of nodes or edges in the source or target version.

0050A first naïve delta representation would be to save all bounds which have changed between V and V′. However, a node insertion changes an average of n/2 bounds yielding ο(n) space complexity. Obviously, this violates the required space complexity bound. Such space complexity is even equal to saving the interval encoding of the target version thoroughly.

0051Our technique for delta storage leverages the fact that each change introduces only a constant number of shifts (translations) of ranges of interval bounds. The key observation is that each atomic update of a tree (node insertion, node deletion, subtree relocation, subtree deletion) can be modeled in the interval bound space by a swap of two ranges of interval bounds, followed by an update of the max value in case of insertion or deletion to adjust the size of the active region. <figref idref="DRAWINGS">FIG. 3</figref> depicts the implementation of the operations using the swapping of two ranges. Subtree deletion is omitted as it is conceputally equal to node deletion. The middle of the figure shows the relocation of the subtree rooted at node C to the right of node E. The hierarchy is shown on the top and its interval encoding is displayed beneath. The relocation is simply accomplished by swapping the bound ranges R₂(all bounds of the subtree C) and R₃(all bounds between the subtree and the target position). This swap constists of translating all bounds in the range R₂by +2 and all bounds in R₃by −4. By storing only these range translations, we achieve ο(c) space complexity. Hereinafter, we call the ranges of bounds which can be translated together simply translation ranges. The left side of the figure shows the insertion of a new node F as rightmost child of node B. The dashed lines of F depict that F is a virtual node prior to the insert. Again, this insertion is accomplished by swapping regions R₂and R₃. Since we are using the active region approach for adding new nodes, the max value of the resulting version is incremented by +2 because a new node was added. Node deletion (and subtree deletion) is similar to insertion: The right of <figref idref="DRAWINGS">FIG. 3</figref> shows the opposite of the left side: Node F is deleted by swapping the regions and reducing the max value.

0052Formally, let swap([a,b],[c,d]) be the function that swaps the interval [a,b] with the interval [c,d] under the preconditions that c=b+1 (the intervals are adjacent and the second one is behind the first one), a≦b̂c≦d (the intervals are well-formed, non-empty intervals). Let insrt([x,y],z) be the method that inserts non-empty interval [x, y] before z under the precondition that z∉[x, y] and x≦y. The function insrt is implemented through a swap:

0000<maths id="MATH-US-00001" num="00001"><math overflow="scroll"><mrow><mrow><mi>insrt</mi><mo></mo><mrow><mo>(</mo><mrow><mrow><mo>[</mo><mrow><mi>x</mi><mo>,</mo><mi>y</mi></mrow><mo>]</mo></mrow><mo>,</mo><mi>z</mi></mrow><mo>)</mo></mrow></mrow><mo>=</mo><mrow><mo>{</mo><mtable><mtr><mtd><mrow><mrow><mi>swap</mi><mo></mo><mrow><mo>(</mo><mrow><mrow><mo>[</mo><mrow><mi>z</mi><mo>,</mo><mrow><mi>x</mi><mo>-</mo><mn>1</mn></mrow></mrow><mo>]</mo></mrow><mo>,</mo><mrow><mo>[</mo><mrow><mi>x</mi><mo>,</mo><mi>y</mi></mrow><mo>]</mo></mrow></mrow><mo>)</mo></mrow></mrow><mo>,</mo></mrow></mtd><mtd><mrow><mrow><mi>if</mi><mo></mo><mstyle><mspace width="0.8em" height="0.8ex" /></mstyle><mo></mo><mi>z</mi></mrow><mo>&lt;</mo><mi>x</mi></mrow></mtd></mtr><mtr><mtd><mrow><mrow><mi>swap</mi><mo></mo><mrow><mo>(</mo><mrow><mrow><mo>[</mo><mrow><mi>x</mi><mo>,</mo><mi>y</mi></mrow><mo>]</mo></mrow><mo>,</mo><mrow><mo>[</mo><mrow><mrow><mi>y</mi><mo>+</mo><mn>1</mn></mrow><mo>,</mo><mrow><mi>z</mi><mo>-</mo><mn>1</mn></mrow></mrow><mo>]</mo></mrow></mrow><mo>)</mo></mrow></mrow><mo>,</mo></mrow></mtd><mtd><mi>otherwise</mi></mtd></mtr></mtable></mrow></mrow></math></maths>

0053Using insrt and the active region approach, implementing all update operations is straightforward:

0054insertBefore(b): insrt([max+1, max+2], b) max:=max+2

0055moveSiblingRangeBefore(v,v′,b): insrt([v.low,v′.high],b)

0056deleteSiblingRange(v,v′): insrt([v.low, v′.high], max+1)

0000max:=max−(v′.high−v.low+1)

0057Since all update operations are now reduced to swap, updating a delta solely relies on efficient implementation of this function, as will be described below. In the example, the insertion is performed by executing swap([6,9], [10,11]), the move by swap([3,6], [7,8]), and the deletion by swap([6,7], [8,11]).

0058We represent version deltas compactly as the ordered set of all translation ranges that were introduced by updates that happened between the source and the target version. We represent the ranges by storing the value of their borders in the source space and in the target space. The value by which the ranges were translated can be inferred by computing the difference between these two numbers. We distinguish between bounds and borders: All values represented by the delta are values of (interval) bounds. A lower/upper border of a translation range is the least/greatest bound that lies in this range Because the translation ranges are densely arranged next to each other, it is sufficient to store only the lower border of the ranges, that is, the value of the lowest bound which lies inside the range. The upper border can be inferred by looking up the lower border of the successive range and substracting one. The highest range (that has no successor) is unbounded, i.e., its upper bound is the positive infinity.

0059<figref idref="DRAWINGS">FIG. 4</figref> shows a graphical representation of how our approach represents a delta. Specifically, the figure shows the delta resulting from the node insertion which is depicted on the left of <figref idref="DRAWINGS">FIG. 3</figref>. The vertical bars represent the lower borders of the translation ranges and the arrows depict to which position these borders are translated. An update introduces at most three new translation ranges: The two ranges that were swapped (R₂and R₃in the example) and the range behind them (R₄in the example). Since only the lower borders of translation ranges are stored, the range before the swapped ones (R₁in the example) does not have to be adjusted; its upper border is adjusted implicitly. We use the notation R(s,t) to denote a translation range which maps from s in the source space to t in the target space. The width of the range is not included in this notation, as it is implicitly defined by the border of the following range in the delta. Thus, the delta depicted in <figref idref="DRAWINGS">FIG. 4</figref> is written as {R(0,0), R(6,8), R(10,6), R(12,12)}.

0060Using this representation, the delta function δ(b) is implemented as follows: Find the translation range with the greatest lower border which is equal to or less than b. Let t be the translation of that range. Then, delta is defined as δ(b)=b+t. For example, the bound <b>7</b> in the source space lies in the yellow translation range (because its lower border <b>6</b> is the greatest which is less or equal to 7), so it is translated by +2, yielding δ(7)=9. Note that this representation also allows to compute δ⁻¹similarly by applying the reverse translation. For example, the bound <b>6</b> in the target space lies in the green translation range. Therefore, δ⁻¹(6)=6−(−4)=10.

0061The representation shown in <figref idref="DRAWINGS">FIG. 4</figref> is a conceptual model instead of a specific data structure. A suitable implementation of this model may satisfy the following conditions:

00621. Because computing δ(b) comprises of finding the greatest lower border of a translation range which is less than or equal to b, the implementation allows finding such border in at most ο(log c) which was defined as the upper bound for the time complexity of δ.

00632. Since the implementation may also be used for the reverse mapping δ⁻¹, it allows execution of this operation in ο(log c).

00643. As shown below, it is possible to build a version delta incrementally by performing range swaps. The implementation provides efficient support for this operation.

0065One implementation uses self-balancing search trees to satisfy conditions <b>1</b> and <b>2</b>, as they allow for ο(log c) query time. Concretely, the implementation comprises two balanced search trees, one for the source and one for the target space. Consequently, we call the trees source tree and target tree. Using these trees enables efficient computation of δ and δ⁻¹. The keys in the search trees are the values for the lower bounds of the translation ranges. The payload in the search tree entries consists of a pointer to the corresponding node in the respective other search tree. The translation value can be omitted, because it can be calculated by following the pointer from the source tree node to the corresponding target tree node and substracting the key in the target space from the key in the source space. <figref idref="DRAWINGS">FIG. 5</figref> shows the source and target tree for the translation ranges from <figref idref="DRAWINGS">FIG. 4</figref> using balanced binary search trees. The dashed arrows represent the pointers between the corresponding nodes in the trees.

0066A pair of source and target tree initially seems to be a memory-consuming data structure due to bidirectional pointers between corresponding nodes. However, the tree can be compressed well: Since one entry in the source tree always corresponds to exactly one entry in the target tree, it is possible to allocate these entries together. Such an entry pair does not need any pointers; instead, implicit address calculation can be used to get from the source entry to the target entry and vice versa. For simplicity, we will still refer to pointers below. However, the pointers are only used for the sake of clarity and can be optimized away.

0067Using the source/target tree representation, the implementation of δ(b) is straightforward: Usual search tree lookup in the source tree is used to find the translation range with the greatest lower border less or equal to b. By following the pointer to the corresponding node in the target tree and lookup up its value there, the translation value is calculated. The implementation of δ⁻¹(b) is equally straightforward: Lookup b in the target tree instead of the source tree and apply the negated translation value.

0068The size of the delta is in O(c) but is also bounded by the size of the hierarchy: The largest delta contains one translation range for each bound of the hierarchy. Repeated updates of a node or subtree (e.g., moving a tree around twice) do not yield extra translation ranges but only update the already existing changes. Therefore, the size of a delta can also be less than O(c) in the case of repeated updates of nodes.

0069We have shown an approach for storing version deltas by representing translation ranges as nodes in two search trees which are linked with each other. This approach allows computing the δ function and its inverse δ⁻¹efficiently. The remaining challenge is to build this data structure efficiently. There are different possible scenarios for building a delta: One scenario (which we call “static”) could be the case that the source and the final target version are available as usual interval encodings and the delta is to be inferred from them. A more dynamic scenario would consist of building the delta incrementally: Whenever a new update is performed on the tree, the resulting translation range swap is added to the data structure. Handling this scenario efficiently is more difficult and requires using specially augumented search trees.

0070In the static scenario we assume that the source and target version for which to build a delta are fully available as interval encodings. This could be the case in applications where a user fetches a whole version from the database, edits it with a third party program (e.g., a graphical tree editor) and then saves the result back to the database creating a new versions. Another use case would be the periodic gathering of snapshots from the web. The operations performed on the tree are not known in this scenario, only the resulting interval encoding is available or is constructed on the fly. Although it might seem easier to infer the delta if the update operations are known, the opposite is actually the case: Inferring a delta from two interval encodings is a straightforward and efficient operation. There is only one requirement: Given a bound value in the interval encoding, it must be possible to infer the node to which this bound value belongs and a matching of nodes must be available; such matching is either implicit if the nodes carry unique identifiers (as in HR hierarchies, our example, and many other use cases) must be used to match nodes in the two versions.

0071The algorithm for inferring the delta from two interval representations I_sand I_tis as follows. Traverse the interval encoding I_sfrom the lowest to highest bound. For each bound b, find the corresponding bound in I_t, i.e. look up the node v to which b belongs and look up the corresponding bound b′ of the node matching v in I_t. Compute the translation t by substracting b′ from b. Store this translation as t′ until the next bound is traversed. Whenever the translation value of adjacent bounds changes, i.e., t≠t′, then insert a new translation range R(b, b′) into the delta. In addition, insert a translation range border for the first bound that is traversed.

0072<figref idref="DRAWINGS">FIG. 6</figref> shows the result of the algorithm comparing a source hierarchy with two target hierarchies. The source hierarchy is shown on the left. The lower and upper bounds belonging to each node are displayed on the left and right of the node, respectively. The middle of the figure shows a target hierarchy where only one update has occurred: The node E was moved behind node C. Below the hierarchy, the resulting bound values are displayed. The first row shows the interval bounds in the source interval encoding I_s. The second row shows the values of the corresponding bounds in I_t. The third row shows the resulting translation. The delta is inferred by the third row from left to right and inserting a new translation range border whenever the value changes (columns with changing value are hightlighted). In addition, as mentioned, a translation range border is always inserted for the first column. So, in the example, the value changes three times: First, it changes from 0 to +2 at bound <b>5</b>, then it changes from +2 to −3 at bound <b>8</b>, and finally, the value changes from −3 to 0 at bound <b>10</b>. An extra border is inserted at bound <b>1</b> which is the first bound to be traversed. Thus, the resulting delta contains four translation ranges, namely {R(1,1), R(5,7), R(8,5), R(10,10)}. The right side of <figref idref="DRAWINGS">FIG. 6</figref> shows a target hierarchy where more changes were introduced. Consequently, there are also more translation ranges (6) in the resulting delta.

0073The previous section introduced a scheme which bulk-builds a delta from two fully materialized interval encodings. This approach has two drawbacks:

00741. It is often undesirable to fully materialize both interval encodings of the source and the target version. The advantage of our approach is that one does not need to materialize these encodings. To leverage this advantage, another technique must be used for updating version deltas.

00752. Rather than introducing single changes at a time, all changes must already be contained in the interval encoding of the target version. Since integer interval encodings are inefficient when it comes to updates, this poses a problem.

0076It would be more appropriate if a version delta could directly be updated efficiently without having to infer any explicit interval encodings. As mentioned above, every atomic update can be modeled by a swap of two consecutive translation ranges. Thus, an efficient update mechanism performs this swap efficiently. The operation swap([a,b],[c,d]) for a delta δ performs the swap in the target space. The bound coordinates a,b,c,d are also in the target space. Conceptually, the operation is implemented as follows:

00771) Insert the translation ranges R(δ⁻¹(a), a), R(δ⁻¹(c), c), and R(δ⁻¹(d+1),d+1). If any of the ranges already exists, do not insert it again. For all translation ranges R(s,t) in δ with t∈[a,b], translate t by the size of [c,d] (i.e., by d−c+1).

00782) For all translation ranges R(s,t) with t∈[c,d], translate t backwards by the size of [a,b].

0079The top of <figref idref="DRAWINGS">FIG. 7</figref> depicts a delta in which the ranges R₁=[2,4] and R₂=[5,14] are to be swapped. The delta already contains nine translation ranges (A-I). The middle of the figure shows the result after performing the first step of the algorithm: C′=R(5,2), which is the lower border of R₁, and E′=R(11,5), which is the lower border of R₂, are inserted. The lower border H=R(20,15) of the range after R₂is already included in the delta, therefore, it can be reused and no new border is inserted. The bottom of the figure shows the delta after performing the second step of the algorithm: The target values of translation ranges that lie in R₁in the target space have been translated by +10 and the target values of those in R₂have been translated by −3.

0080The implementation of swap reveals two performance bottlenecks: First, the target value of all bounds in R₁(marked orange in the figure) and R₂(marked green) has to be adjusted. On average, this results in n/2 adjustments which yields an infeasible linear runtime if done naïvely. Next, since the target values are also keys in a search tree, they cannot be simply adjusted. Instead, their swapped position is updated in the search tree as well by also swapping the nodes in the search tree. Otherwise, the tree would no longer be ordered by key and thus would malfunction. Again, n/2 nodes in the search tree need to be swapped. If this is done naïvely by deletion and reinsertion, it would even yield ο(n log n) time complexity.

0081To allow efficient updates of an interval delta in ο(log n), the search tree which models the target space has to be augumented to allow adjusting multiple keys at once and swapping ranges of search tree nodes efficiently.

0082The efficient swapping of nodes can be accomplished by adding the split and join functionality to the self-balancing search tree: The split(T,k) function splits the search tree T before a key k, resulting in a tree that holds all keys &lt;k and one that holds all keys ≧k. Both resulting trees are appropriately balanced. Given two search trees T₁and T₂where all keys in T₂are greater than all keys in T₁, the join(T₁,T₂) function concatenates the trees resulting in a new search tree which contains all keys of the trees and must be appropriately balanced again. Although both functions are quite uncommon since they are not needed by RDBMSs, an efficient ο(log n) implementation exists for all common self-balancing search trees. We can swap two ranges of search tree nodes by first splitting the tree at the borders of these ranges and then joining the resulting trees in a different order. One can imagine this operation as simply cutting the tree into smaller trees representing the different ranges and then gluing these trees together in the desired order. Such a swap consists of three splits and three joins and is therefore in ο(log n) matching the desired time complexity bound.

0083The second requirement is satisfied by replacing the ordinary search tree with a slightly adapted implementation which we call accumulation tree. An accumulation tree is a search tree in which each node only stores a part of its own key. The real key of a node v is obtained by adding (accumulating) all values on the path from v to the root. Since a search tree already traverses this path during the key lookup, the accumulation of the key of v is cheap. <figref idref="DRAWINGS">FIG. 8</figref> shows the delta from <figref idref="DRAWINGS">FIG. 5</figref> with an accumulation tree used as target tree. The resulting accumulated values (which are equal to the values of the original target tree in <figref idref="DRAWINGS">FIG. 5</figref>) are shown in parenthesis below the nodes. For example, the rightmost node has a value of 12. This value is obtained by accumulating all values (6, 2, 4) on the path from the root to this node.

0084The idea behind the accumulation tree introduces an important property: All keys in a subtree rooted at a node v can be translated by simply adjusting the value of v, resulting in ο(1) time complexity instead of ο(n) (with n being the size of the subtree). However, the tree introduces a small maintenance overhead: Whenever performing rotations to restore the balance of the tree, the values of the two rotated nodes and the value of the root of the middle sub-tree have to be adjusted. For left rotations, this is the left child of the right child of the rotated node. For right rotations, this is the right child of the left child of the rotated node. Otherwise, the rotation would alter the accumulated values.

0085<figref idref="DRAWINGS">FIG. 9</figref> depicts the rules for updating the values after a left rotation. If the node that is rotated has the value x, its right child has value y and the left child of the right child has value z, then the three nodes must receive the values y+x, −y, and z+y, respectively. For example, the root of the subtree in the middle has x+y+z as accumulated value before the rotation. Afterwards, it still has (y+x)+(−y)+(z+y)=x+y+z. All other nodes behave similarly, so the rotation does not alter any accumulated value. Performing a right rotation is similar.

0086Using the accumulation tree in combination with the split and join operations, we are able to achieve the swap of two translation ranges in ο(log n). The steps of the algorithm performing the swap from <figref idref="DRAWINGS">FIG. 7</figref> are depicted in <figref idref="DRAWINGS">FIG. 10</figref>. On the top left of the figure, the target tree (without using accumulation) of the model is shown. The source tree is completely omitted in the figure, as it is not altered by a swap, except that the new translation ranges are entered at appropriate positions. The lower left part of the figure (Step 1) shows the target tree using from the top, but now using the accumulation tree. For example, the node I has a value of 23. This value is represented in the accumulation tree by the accumulated values 5,3,5,10. The dashed lines represent the positions where the tree is split. Step 2 of the figure shows the resulting trees after the splits are performed. Note that the split also rebalances the resulting trees. The next step (3) is to apply the translations to the two ranges. The accumulation tree allows this operation by simply adjusting the value in the root. The root of R₁(F) is translated by −3 and the root of R₂(E) is translated by +10. The trees are then joined to yield the resulting tree. Of course, the tree for R₂is joined to the left of R₁to achieve the swap, so the trees are joined in the order C, F, E, G. The resulting (rebalanced) tree is shown on the right of the figure. Thus, in conclusion, a swap of two adjacent translation ranges is simply performed by first splitting the tree, applying the transformations to the roots of the resulting trees and then re-joining the trees in a different order. Since the time complexity of split and join is in ο(log n) and the complexity of the translation in the accumulation tree is in ο(1), the resulting time complexity is ο(log n). In contrast to the linear time complexity of the basic data structure, the accumulation tree thus allows for efficient updates.

0087The previous section has introduced an efficient storage technique for interval deltas. Such delta maps interval bound from a version V to another version V′ and vice versa. Now assume a large version history with n versions V₀, . . . ,V_n-1. To be able to answer queries for an arbitrary version V_i, one or more deltas must exist which eventually lead from a base version to version V_i.

0088Without loss of generality, we will hereinafter assume a linear version history (without any branches) with only one base version which is the eldest version V₀. The version indices are sorted by the age of the version, so V_iis the version right before V_i+1and right after V_i−1. We define the size of a delta, written as |δ|, by the number of versions captured by it. For example, the delta δ_V_{<sub2>20</sub2>}_→V_{<sub2>30 </sub2>}would have the size 10, because it covers all changes introduced in the ten versions versions V₂₁, . . . , V₃₀. If a constant number of changes per version is assumed, the size of the delta is proportional to its memory consumption.

0089The interval bounds of each node are known for the base version V₀. Since we represent a versioned hierarchy by a table and the DeltaNI index, we store the interval bounds of V₀in the table itself by adding a low and high column. Deltas are used to transform these bounds into each other version. The bounds in V₀also pose durable identifiers for all the nodes, since they will not change. By maintaining an index over the low column, one can access a node by its low bound in V₀. To retrieve the node belonging to a bound b in another version, we need to apply inverse deltas to transform b back to V₀; then we can lookup the corresponding node via the index.

0090Let δ₁, . . . , δ_mbe a sequence of deltas where each delta δ_imaps from a version to the version of the subsequent delta δ_i+1. If the first delta maps δ₁from V_sand the last delta maps to V_t, then we can retrieve the bound b_tin V_tfor a bound b_sin V_sby applying all deltas in the sequence:

0000<br /><i>b</i>_t=δ_m(δ_m−1( . . . δ₂(δ₁(<i>b</i>_s)) . . . ))

0091By applying the deltas reversely in the opposite order, we can also map back from V_tto V_s.

0092The time complexity of such query is in ο(m log c), where c is the maximum of changes contained in any of the deltas in the sequence. So, for fastest query times, a sequence length of one would be best. This, however, implies that a delta from a base version to each other version must exist. Since we assume a linear version history as the worst case, a change introduced in a version V_iwill also be stored in the interval deltas between the base version and all versions which are more recent than V_i. When assuming a constant number of changes per version, maintaining deltas from the base version to each other version would require ο(n²) space in the worst and best case, because each change is contained in n/2 deltas on average. This is not feasible for hierarchies with millions of versions. Another extreme would be to save only the deltas from version V_ito V_i+1. Assuming a constant number of changes per version would yield ο(n) space complexity, because each change is only stored in the delta for the version in which it was introduced. This is the strategy with the least space consumption. However, a query in version V_iwould then require i delta applications since all deltas of versions older than V_ihave to be applied one by one. On average, this yields ο(n) query complexity which is infeasible for large hierarchies, as well.

0093We achieve a good space/time trade-off by enforcing an exponential distribution of the delta sizes. That is, some few deltas cover huge version ranges while most deltas cover only a few versions. The large deltas can be used to get “near” the target version quickly. There, the small deltas are used to get exactly to the target version. This approach is comparable to the one of skip lists or to the finger tables in the peer-to-peer protocol.

0094Our approach uses the number of trailings zeros in the binary representation of the id of a version to determine the size of the delta leading to this version. Precisely, given a version V_i, the size of the delta δ which has V_ias target version is calculated as |δ|=2^tz(i), where tz(x) is the number of trailing zeros in the binary representation of x. For example, version 27 has the binary representation 11011₂. Since this binary string has no trailing zeros, this version will be represented by the delta δ_V_{<sub2>26</sub2>}_→V_{<sub2>27 </sub2>}which has a size of 1. In contrast, version 36 corresponds to the binary string 100100 which has two trailing zeros. This results in the delta δ_V_{<sub2>32</sub2>}_→V_{<sub2>36 </sub2>}of size 2²=4. <figref idref="DRAWINGS">FIG. 11</figref> depicts the size of the first ten interval deltas of a version history.

0095To query a version V_iusing this technique, one has to start at the base version and execute “hops” which become smaller and smaller. The arrow in <figref idref="DRAWINGS">FIG. 11</figref> shows how a query for version 7 is processed. The algorithm for finding the hops for version V_isimply consists of scanning the bit positions j of the binary representation of i from most significant bit to least significant bit. Whenever a one bit is found at position j, take one hop. The target version is i with all less significant bits than j zeroed out. For example, a query in version i=19=10011₂would be processed as follows: The highest one bit is j=4 (j is counted from least to most significant bit, starting with zero for the least significant one), so the first hop is to version 10000₂=16. The next one is at j=1, resulting in the hop to 10010₂=18. The final hop for the last one bit at j=1 is 10011₂which reaches the target version 19. The resulting deltas to be applied are V₀→V₁₆, V₁₆→V₁₈, and V₁₈→V₁₉.

0096Since the algorithm takes one hop per one-bit of the version id i and version id bit lengths are logarithmic in the number of versions, the number of deltas to be applied to reach a version V_iis ┌log₂(i)┐ in the worst case (when the version id consists only of one-bits) and 1 in the best case (when the version id is a power of two). When maintaining a version history of n versions with n being a power of two, each bit of a randomly chosen version id i is one or zero with the same probability, so the algorithm applies log₂(n)/2 deltas on average.

0097A change introduced in version V_iis contained in the version delta for V_iand all version deltas of all higher versions where i is a power of two. For example, a change introduced in version 7 in contained in the deltas V₀→V₇, V₀→V₈, V₀→V₁₆, V₀→V₃₂, and so on. Obviously, for a version history of n versions, there are logarithmically many versions which are a power of two, so each change is contained in at most 1+┌log₂n┐ versions. Since one change needs a constant amount of space, a version history with n versions and constant number of changes per version can be stored using ο(n log n) space (ο(n) changes in total, each being stored in ο(log n) versions).

0098As already shown, applying a delta of size c to a single bound has a time complexity of ο(log c). However, the more relevant question is the time complexity of retrieving the value of a bound b in a version V_iwhich needs to apply more than one delta. In the worst case, when the binary representation of i has only ones in it, the algorithm must apply log₂i deltas. The number of versions covered by each of these deltas is exponentially distributed: The first covers i/2 versions, the second i/4 versions, and so on until the last one only covers a single version. If we assume a constantly bounded number of changes c per version (which is a reasonable assumption if new versions are created regularly), the number of changes contained in the deltas is proportional to the number of versions v covered by it and thus the complexity of applying this delta is ο(log v). Consequently, the complexity of applying all the deltas is ο(Σ_j=0^log^{<sub2>2</sub2>}ⁱlog 2^j)=(Σ_j=0^log^{<sub2>2</sub2>}ⁱj)=ο(log²i) in the worst case. In the best case, the version number is a power of two and thus only one delta has to be applied, yielding ο(log i).

0099During the generation of the exponentially sized deltas, smaller deltas have to be merged to yield larger ones. For example, the delta V₀→V₈is to built by first merging the deltas V₀→V₄, V₄→V₆, and V₆→V₇, which yields the delta V₀→V₇. Now, there are two equally applicable strategies: One strategy is to apply the incoming changes for V₈directly to the delta V₀→V₇, yielding the delta V₀→V₈without further merges. Another strategy is to gather the changes for V₈in a small delta V₇→V₈and finally merge V₀→V₇with V₇→V₈to yield the final delta V₀→V₈. Regardless of the strategy used, an operation for merging two deltas is required.

0100Let, δ_V→V′ and δ_V′→V″ be two deltas which are connected via the version V′, i.e., V′ is the source of the one and the target of the other delta. We define the operation merge(δ_V→V′, δ_V′→V″) which merges the changes in the two deltas yielding the delta δ_V→V″. The resulting delta function must be the composition δ_V→V′∘δ_V′→V″, i.e.:

0000<br />∀<i>b∈N·δ</i>_V→V″(<i>b</i>)=δ_V′→V″(δ_V→V′(<i>b</i>))

0101The merge(δ₁, δ₂) function can be implemented as follows: Start with an empty delta δ. For each translation range R(s,t) in δ₁, compute t′=δ₂(t) and insert R(s,t′) into δ. Next, for each translation range R(s,t) in δ₂, compute s′=δ₁⁻¹(s). If no translation rule with source value s′ exists in δ, then add R(s′, t) to δ.

0102The implementation basically adjusts all translation ranges in the two deltas to incorporate the changes of the other delta as well. Ranges in the prior delta δ₁need their target values adjusted by δ₂, since the resulting delta maps to the target space of δ₂. The source values of the ranges in δ₂need to be adjusted “backwards” by the inverse of δ₁, because the resulting delta maps from the source space of δ₁. Since each range in δ₁and δ₂adds at most one translation range to the resulting delta, the delta size |δ| is at most |δ₁|+|δ₂|. However, when the ranges of δ₂are processed, they are only added if no delta with the same source value already exists. Thus, the resulting delta size may be smaller than |δ₁|+|δ₂|. The omission of a range is executed if both versions transform the range. For example, if δ₁moves a node X and δ₂moves the same node again, then they will both contain a translation range starting at the lower bound of X. The resulting delta will only contain one rule for this lower bound.

0103<figref idref="DRAWINGS">FIG. 12</figref> shows an example for a merge. The source tree version V₀is shown on top. In version V₁(left), the subtree B was moved below F. In V₂(right), node E was moved below B. The deltas δ₁(V₀→V₁) and δ₂(V₁→V₂) are displayed below the respective tree. A merge of these deltas results in the delta V₀→V₂which is shown on the bottom of the figure. The translation ranges in the deltas are named with the letters a to h. The letters in the resulting delta show from which translation ranges the ranges in this delta originate. For example, the leftmost translation rule R(0,0) is contained in both deltas as a and e, respectively. Thus, the two ranges are merged into one range a/e. Another example is the range c/f. The first delta has the rule c which is R(8,2). When applying δ₂(2), the resulting target value is 8 (2 lies in range f which is translated by +6), so the resulting rule is R(8,8) which is c/f. The second delta contains the rule f which is R(2,8). The resulting source value for this rule is δ₁⁻¹(2)=8 (2 in the target space of δ₁lies in rule c which has a translation of −6, so the inverse translation is +6). Since R(8,8) already exists, no further range is added. From this example: If a bound is first translated from 8 to 2 and then translated back from 2 to 8, then both translations can be combined to R(8,8).

0104Since each translation range in the deltas has to be processed (linear) and for each range, a delta must be computed (log) and a range must be inserted (log), the resulting time complexity of the merge operation is ο(n log n) where n is the maximum of |δ| and |δ₂|.

0105The more versions are to be covered by a delta, the more merges have to be performed to build the delta. The number of deltas to be merged for a delta that covers v versions is log v. Thus, building a large delta pack could stall the database for a few seconds, which is infeasible. This problem can be mitigated by merging deltas ahead of time in the background. For example, the delta V₀→V₆₄requires the merge of the deltas V₀→V₃₂, V₃₂→V₄₈, V₄₈→V₅₆, V₅₆→V₆₀, V₆₀→V₆₂, V₆₂→V₆₃and V₆₃→V₆₄. Once version V₄₈is built, the deltas V₀→V₃₂and V₃₂→V₄₈can already be merged generating delta V₀→V₄₈. Once version V₅₆is built, this delta can be merged with V₄₈→V₅₆to build the delta V₀→V₅₆, and so on. Using this strategy, only the latest two deltas need to be merged at a time, drastically reducing the time of a single delta build and thus preventing longer stalls. However, this strategy requires to maintain the intermediate delta merges like V₀→V₅₆and thus increases memory consumption.

0106Consider a version history which currently contains n versions. It is desirable to achieve the best case query complexity when querying the most recent version V_n. Not only is it much more probable that queries will be executed in this version, but the deltas to this version also need to be calculated when performing updates. To achieve this, we not only maintain the exponentially distributed deltas but also an additional large delta δ_0→nwhich can be used to directly reach the most recent version. The additional maintenance of this delta is only a neglectable constant space overhead in comparison to the exponentially distributed deltas.

0107<figref idref="DRAWINGS">FIG. 13</figref> is a block diagram of system <b>1300</b> according to some embodiments. System <b>1300</b> may comprise a general-purpose computing system and may execute program code to perform any of the processes described herein. System <b>1300</b> may comprise an implementation of database <b>100</b> according to some embodiments. System <b>1300</b> may include other unshown elements according to some embodiments.

0108System <b>1300</b> includes processor <b>1310</b> operatively coupled to communication device <b>1320</b>, data storage device <b>1330</b>, one or more input devices <b>1340</b>, one or more output devices <b>1350</b> and memory <b>1360</b>. Communication device <b>1320</b> may facilitate communication with external devices, such as a reporting client, or a data storage device. Input device(s) <b>1340</b> may comprise, for example, a keyboard, a keypad, a mouse or other pointing device, a microphone, knob or a switch, an infra-red (IR) port, a docking station, and/or a touch screen. Input device(s) <b>1340</b> may be used, for example, to enter information into apparatus <b>1300</b>. Output device(s) <b>1350</b> may comprise, for example, a display (e.g., a display screen) a speaker, and/or a printer.

0109Data storage device <b>1330</b> may comprise any appropriate persistent storage device, including combinations of magnetic storage devices (e.g., magnetic tape, hard disk drives and flash memory), optical storage devices, Read Only Memory (ROM) devices, etc., while memory <b>1360</b> may comprise Random Access Memory (RAM).

0110Database engine <b>1332</b> may comprise program code executed by processor <b>1310</b> to cause apparatus <b>1300</b> to perform any one or more of the processes described herein. Embodiments are not limited to execution of these processes by a single apparatus. interval deltas <b>1334</b>, interval encodings <b>1336</b> and versioned hierarchical data <b>1338</b> may be implemented as described above. As also described above, database <b>100</b> may be implemented using volatile memory such as memory <b>1360</b>. Data storage device <b>1330</b> may also store data and other program code for providing additional functionality and/or which are necessary for operation of system <b>1300</b>, such as device drivers, operating system files, etc.

0111The foregoing diagrams represent logical architectures for describing processes according to some embodiments, and actual implementations may include more or different components arranged in other manners. Other topologies may be used in conjunction with other embodiments. Moreover, each system described herein may be implemented by any number of devices in communication via any number of other public and/or private networks. Two or more of such computing devices may be located remote from one another and may communicate with one another via any known manner of network(s) and/or a dedicated connection. Each device may comprise any number of hardware and/or software elements suitable to provide the functions described herein as well as any other functions. For example, any computing device used in an implementation of system <b>100</b> may include a processor to execute program code such that the computing device operates as described herein.

0112All systems and processes discussed herein may be embodied in program code stored on one or more non-transitory computer-readable media. Such media may include, for example, a floppy disk, a CD-ROM, a DVD-ROM, a Flash drive, magnetic tape, and solid state Random Access Memory (RAM) or Read Only Memory (ROM) storage units. Embodiments are therefore not limited to any specific combination of hardware and software.

0113Embodiments described herein are solely for the purpose of illustration. Those skilled in the art will recognize other embodiments may be practiced with modifications and alterations to that described above.