1.5.0

Fixed

• RelANNIS version 3.3 files with segmentation might also have a missing "span" column.
  In case the "span" column is null, always attempt to reconstruct the actual value from
  the corresponding node annotation instead of failing directly.

Changed

• Avoid unnecessary compacting of disk tables when collecting graph updates during import.
  This speeds up both the GraphML and the relANNIS importer and can also reduce the
  used main memory during import.
• Use release optimization of some of the performance sensitive crates even for debug builds.
  This allows faster builds and debugging of our own code, while balancing performance.

1.4.1

Fixed

• Avoid unnecessary memory allocation when checking if a node has outgoing edges in
  adjacency lists. This improves search for tokens because the Coverage components
  are typically adjacency lists, and we need to make sure the token nodes don't
  have any outgoing edges.
• Fixed miscalculation of whitespace string capacity which could lead to
  memory allocation failed error.

1.4.0

Added

• Added clear() method to the WriteableGraphStorage trait.

Fixed

• Limit the used main memory cache per DiskTable by only using a disk block cache for the C1 table.
  Since we use a lot of disk-based maps during import of relANNIS files, the previous behavior could
  add up to > 1GB easily, wich amongst other issues caused #205 to happen.
  With this change, during relANNIS import the main memory usage should be limited to be less than 4GB,
  which seams more reasonable than the previous 20+GB
• Reduce memory footprint during import when corpus contains a lot of escaped strings (as in #205)
• Avoid creating small fragmented main memory when importing corpora from relANNIS to help to fix #205

Changed

• Improved overall import speed of relANNIS corpora and when applying graph updates

1.3.0

Added

• The webservice endpoint /search/node-descriptions now returns
  wether a node in the query is optional or not.

1.2.2

Fixed

• Queries with optional nodes with a smaller index than the last non-optional node could fail.
  If the execution nodes re-order the match result vector internally, the query node index is
  used to define the mapping. Unfortunately the largest index could be larger than the size of mappings,
  which used to be used to create the output vector. By allowing empty elements in the output vector
  and using the maximum value, we can still map the results properly.

1.2.1

Fixed

• Don't allow optional operands for non-negated operators

1.2.0

Added

• Added generic operator negation without existence assumption,
  if only one side of the negated operator is optional (#187).

1.1.0

Added

• Added generic operator negation with existance assumption by adding ! before the binary operator (#186)

Changed

• Compile releases on Ubuntu 18.04 instead of 16.04, which means the minimal GLIBC version is 2.27
• Updated dependencies
• Improved compile time by disabling some dependency features.
  This also removes some optional features from the command line parser
  (used in webservice and CLI binaries).
• Don't use RIDGES corpus in search tests and fail search tests when corpus does not exist.

Fixed

• Use the correct set-disk-based on command in the documentation for the CLI
• Optimize node annotation storage and graph implementations when importing GraphML files

1.0.2

This release should not have any functional changes and only fixes issues in the GitHub Actions pipeline.

Fixed

• Fix issue when deploying release artifacts on GitHub

1.0.1

Fixed

• Assume that the annis::node_name annotation is unique when estimating match size.
  This should improve e.g. subgraph-queries, where the intermediate result sizes are now better estimated.