Span started as a single tool to convert Crossref API data into a VuFind/SOLR format as used in finc. An intermediate representation for article metadata is used for normalizing various input formats. Go was choosen as the implementation language because it is easy to deploy and has concurrency support built into the language. A basic scatter-gather design allowed to process millions of records fast.
While span has a few independent tools (like fetching or compacting crossref feeds), it is mostly used inside siskin, a set of tasks to build an aggregated index.
$ go install github.com/miku/span/cmd/...@latest
Span has frequent releases, although not all versions will be packaged as deb or rpm.
Initial import Tue Feb 3 19:11:08 2015, a single span command. In March
2015, span-import and span-export appeared. There were some rudimentary
commands for dealing with holding files of various formats. In early 2016, a
licensing tool was briefly named span-label before becoming span-tag. In
Summer 2016, span-check, span-deduplicate, span-redact were added, later
a first man-page followed. In Summer 2017, span-deduplicate was gone, the
doi-based deduplication was split up between the blunt, but fast
groupcover and the generic
span-update-labels. A new span-oa-filter helped to mark open-access
records. In Winter 2017, a span-freeze was added to allow for fixed
configuration across dozens of files. The span-crossref-snapshot tool
replaced a sequence of luigi tasks responsible for creating a snapshot of
crossref data (the process has been summarized in a
comment).
In Summer 2018, three new tools were added: span-compare for generating index
diffs for index update tickets, span-review for generating reports based on
SOLR queries and span-webhookd for triggering index reviews and ticket
updates through GitLab. During the development, new input and output formats
have been added. The parallel processing of records has been streamlined with
the help of a small library called
parallel. Since Winter 2017, the
zek struct generator takes care of the initial
screening of sources serialized as XML - making the process of mapping new data
sources easier.
Since about 2018 (0.1.211), the span tools have seen mostly small fixes and
additions. Notable, since 2021, the previous scripts used to fetch daily
metadata updates from crossref has been put into a
standalone tool, span-crossref-sync, which merely adds some retry logic and
consistent file naming to the API harvest. In 2024, span-webhookd,
span-check, span-review, span-tagger are gone.
See: manual source.
In the best case no complete processing of the data should take more than two hours or run slower than 20000 records/s. The most expensive part currently seems to be the JSON serialization, but we keep JSON for now for the sake of readability. Experiments with faster JSON serializers and msgpack have been encouraging, a faster serialization should be the next measure to improve performance.
Most tools that work on lines will try to use as many workers as CPU cores.
Except for span-tag - which needs to keep all holdings data in memory - all
tools work well in a low-memory environment.
More cores can help (but returns may diminsh): On a 64 core 2021
Xeon,
we find that e.g. span-export can process (decompression, deserialization,
conversion, serialization, compression) on average 130000 JSON documents/s. The
final pipeline stage (from normalized data to deduplicated and indexable data)
seems to take about three hours.
The span tools are used in various tasks in siskin (which contains all orchestration code). All span tools work fine standalone, and most will accept input from stdin as well, allowing for one-off things like:
$ metha-cat http://oai.web | span-import -i name | span-tag -c amsl | span-export | solrbulk