Move the build into a multi-stage Dockerfile (ubuntu:25.04 + gcc-15) so the
benchmark works on any Ubuntu host without polluting it with Hyrise's
toolchain. The runtime image only carries hyriseServer, libhyrise_impl.so,
libjemalloc.so, and a small set of shared-library deps (~250 MB).

Build args:
- HYRISE_REF (default master) — pin a Hyrise revision
- NO_LTO (default FALSE) — toggle LTO for faster development builds

Loading: drop create.sql and use hits.csv.json next to the data file as the
schema source. CREATE TABLE followed by COPY trips a Hyrise assertion
("set_immutable() should not be called on an empty chunk", chunk.cpp:125)
because COPY tries to seal the empty chunk left by CREATE TABLE; letting
COPY auto-create the table from the CSV meta avoids the issue.

run.sh: detect failed queries via psql's exit code rather than grepping the
output, so errors like "Invalid input error: Could not resolve function
'LENGTH'" are recorded as null. Hyrise lacks LENGTH, REGEXP_REPLACE,
DATE_TRUNC, and OFFSET, so 7 queries (Q28, Q29, Q39-Q43) are reported as
null; the remaining 36 succeed.

Tested locally on arm64: docker build produced a working image, hyriseServer
accepts psql connections, COPY loads a 1000-row sample, and run.sh produces
the expected 43 lines of [t1,t2,t3] output with nulls in the right slots.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Split the old monolithic benchmark.sh + run.sh into the new layout:
install / start / stop / check / load / query / data-size, with
benchmark.sh now a thin shim that exec's lib/benchmark-common.sh.

BENCH_DURABLE=no because Hyrise has no on-disk persistence — the
driver re-runs ./load on every cold cycle and rolls that wall-clock
into the cold-try timing. The load script stages hits.csv (delivered
by download-hits-csv into cwd) under data/ next to hits.csv.json,
which the container exposes read-only at /data; the move is
idempotent so subsequent cold-cycle reloads reuse the staged file.

query keys off psql's exit code with ON_ERROR_STOP=1, so the seven
queries Hyrise can't run (Q28, Q29, Q39–Q43 — LENGTH, REGEXP_REPLACE,
DATE_TRUNC, OFFSET) propagate non-zero to the driver and get recorded
as null instead of stealing the Time line psql still prints for the
\timing meta-command.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Bring in lib/benchmark-common.sh and lib/download-hits-* — the
hyrise/benchmark.sh shim added in f94c8af exec's ../lib/benchmark-common.sh,
but the branch had never been rebased onto main where lib/ lives, so the
shim died with "No such file or directory" on every cloud-init run.

`libhyrise_impl.so` links `libboost_container.so.1.83.0` and
`libboost_date_time.so.1.83.0` dynamically (visible in the CI link
command), but the runtime stage only installed boost-system and
boost-thread, which do not depend on them — `hyriseServer` would fail
to start with a missing shared library once the build succeeds.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…shot

The monolithic `COPY hits FROM '/data/hits.csv'` cannot work at full
scale: Hyrise's CSV import reads the whole file into a single
std::string and materializes every column as unencoded ValueSegments
before dictionary-encoding them — roughly 200 GB peak for the 75 GB
hits.csv. `COPY ... FROM` into an existing table appends chunks, so
./load now splits the file into 5M-row pieces and imports them
sequentially, capping the transient overhead at one piece on top of the
growing encoded table.

The driver (BENCH_DURABLE=no) re-runs ./load on all 43 cold cycles, and
re-parsing the CSV each time would blow any time budget. The first load
therefore exports the table to Hyrise's binary format, which preserves
the encoded segments and streams chunk-by-chunk on import; every reload
imports that snapshot — the closest Hyrise equivalent of a durable
system re-reading its own on-disk format after a restart. The export
must go through `hyriseConsole`'s `export` command: server-side SQL
`COPY TO` wraps the table in an MVCC `Validate` operator and
BinaryWriter materializes the resulting ReferenceSegments as unencoded
value segments, so the snapshot would be raw-sized and would re-import
unencoded. The console is a separate process with its own catalog, so
it performs the piecewise CSV import itself and the server only ever
imports the snapshot (regenerating full-table statistics on add_table).

The runtime image now ships `hyriseConsole` (plus libreadline/
libncurses), and the /data bind mount is writable for the snapshot.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>