Jd is Jsoftware's high-performance columnar RDBMS, written in C with a deep J integration. Non-commercial use is free; a non-commercial key is auto-installed on first run.

This entry uses Jd's native reads query language (SQL-ish but with J syntax) rather than translating to ANSI SQL — Jd accepts SQL keywords in a different order (reads <select> from <table> where <where> order by <order>) and uses by inside reads for GROUP BY. The queries.sql file holds J expressions that wrap jd 'reads …' calls plus J operators for things Jd's query layer doesn't ship (LIMIT, DISTINCT).

./install:

1. Clones jsoftware/jsource@build96 and uses jlibrary/ as the J installation root, then overlays the platform-specific binary (jconsole, libj.so, libtsdll.so, libgmp.so) from the same tag's release zip (l64.zip on x86_64, rpi64.zip on aarch64). The release zip ships binaries only and won't run without jlibrary/'s standard library.
2. Installs a small /usr/local/bin/ijconsole wrapper that re-execs the real jconsole under faketime '2026-05-10 00:00:00'. Why: Jd's bundled jdkey.txt is an evaluation key Jsoftware refreshes periodically, and the copy in jsoftware/data_jd expired 2026-05-16. Until upstream pushes a new key (tracked in the data_jd repo as jdkey.txt), every jconsole invocation needs to see a date before the expiry or jdlicense returns _2 ("eval key") and jd.ijs:147 asserts out. Backdating with faketime is the cheapest workaround that keeps the rest of Jd intact.
3. Uses J's package manager (pacman / jpkg) to install the data/jd addon and its J-side dependencies (api/curl, ide/jhs, arc/lz4, general/misc, data/jfiles, data/jmf, net/jcs, net/socket, web/gethttp, convert/json, convert/pjson).

./load ingests hits.csv via Jd's CSV loader with an explicit column schema instead of csvload_jd_'s auto-inference. The default flow types every string column by sampling the first 5000 rows and then runs csvscan to widen any column it inferred as byte to the full-file max width. ClickBench has very sparse text columns (e.g. OpenstatServiceName, SocialNetwork) that look empty in the 5000-row sample → typed as byte, then later widened to hundreds of chars × 100 M rows. With ~30 such columns the splayed table grew past 500 GB during the load and segfaulted. Declaring text columns as varbyte (variable-length, per-row offset + concatenated data) keeps storage proportional to actual string content. The script writes a hand-rolled hits.cdefs file into the csvload jdcsv folder, then calls csvrd directly, skipping csvcdefs (auto-type) and csvscan (byte-width widening).

Schema choices:

• int (8-byte signed) for every numeric column. Jd's int1 / int2 / int4 leave per-row data as n,x char matrices, and the <> predicate then sees a shape-2 column vs a shape-0 scalar, so we use the flat 8-byte JINT form everywhere.
• varbyte for TEXT / VARCHAR / CHAR.
• edate for EventDate, edatetime for the three TIMESTAMP columns. Both are 8-byte epoch-nanos and Jd's csv loader parses iso8601 from iso8601-char mode (CSV format is YYYY-MM-DD / YYYY-MM-DD HH:MM:SS).

The loader writes per-column files to a dedicated database under ~/j9.6-user/temp/jd/csvload/; that's the database ./query opens.

./query reads a J expression from stdin and evaluates it via ijconsole query.ijs. The query.ijs script opens the csvload database, times the eval, prints the result to stdout, and emits the runtime in fractional seconds to file id 5 (stderr).

The translations stay close to the SQL semantics but diverge in a few places:

• LIMIT n isn't a reads keyword — we use J's n {. after the query (e.g. 10 {. jd '...').
• LIMIT n OFFSET m uses n {. m }. jd '...'.
• COUNT(DISTINCT col) uses J's # ~. (count of unique items) after pulling the column with jd 'reads col from t'.
• min / avg on varbyte: Jd's aggregators are numeric-only, so Q23's MIN(URL) / MIN(Title) become first URL / first Title (any value from each group, semantically ANY_VALUE).
• Q28 (AVG(LENGTH(URL))), Q29 (REGEXP_REPLACE), and Q43 (DATE_TRUNC('minute', ...)) use facilities not in Jd's reads language; they currently return the literal 'null' and the benchmark driver records them as missing. They could be expressed with a J-side computed column — contributions welcome.
• order by requires the column in select: Jd's parser rejects reads SearchPhrase from hits order by EventTime because the order key isn't projected. Q25 / Q27 are rewritten to project EventTime,SearchPhrase (timing unaffected; only the printed output has one extra column).
• COUNT(DISTINCT col): outside reads, J's # ~. ; }. jd '…' (count of unique, after dropping the header row). The }. drops the header box so the unique scan only sees the data values.

J / Jd is single-threaded by default. Jd's columnar layout makes single-column scans fast; cross-column where-then-aggregate paths are also vectorised in the C core. There is no daemon — each query call cold-starts ijconsole, loads the database (mostly memory-mapped columns), and runs.