thptqg/docs/data-pipeline.md

Data Pipeline

From raw Excel files to a compressed SQLite file the browser can load.

One Go binary (parser/) builds every dataset. What differs per dataset is parse rules only — sheet strategy, column layout, validation guards — declared in parser/configs/<id>.yml. The table shape, the INSERT and the subject regexes are canonical and live in parser/internal/schema/schema.go.

Sources

id	Files	Origin	Host live?
2016	4 .xls + 115 .xlsx	dtnt.bacninh.edu.vn aggregator article, 119 exam clusters	yes
2017	63 .xls	baotintuc.vn article, files on its CDN	yes

Full article URLs are below.

Crawling lives in crawler/, a separate Go module. It is never part of the build — the source files are committed, so a crawl only refreshes them. Both runs are idempotent: files already present are skipped.

go -C crawler run ./cmd/crawl 2016     # sources/source_2016.go
go -C crawler run ./cmd/crawl 2017     # sources/source_2017.go
go -C crawler run ./cmd/crawl 2017 --list   # list only, download nothing

2017 comes from the article https://baotintuc.vn/tuyen-sinh/tra-cuu-diem-thi-thpt-2017-cua-63-tinh-thanh-pho-tren-baotintucvn-20170706073512672.htm and its CDN is still serving the files.

2016 comes from the aggregator article https://dtnt.bacninh.edu.vn/tin-tuc/tin-tuc-su-kien/cong-bo-diem-thi-thptqg-2016-toan-bo-120-cum-thi-da-co-diem.html, which lists one spreadsheet per exam cluster. A full crawl against it has been run successfully and reproduces data/2016/, so the dataset is recoverable from source like 2017 is.

That host is not reachable from every network, though: it resolves to a Vietnamese address that times out from at least some hosts abroad, in which case the crawl stops with a connection error before downloading anything. That is a connectivity problem, not a missing dataset — retry from a network that can reach the host.

How a source is defined

A source carries no link list. It names the article that published the files and says how to name what it finds there:

field	meaning
Article	the page to read links from
Exts	which file extensions to pick out of it
WantFiles	how many links to expect — fewer aborts the crawl
Dest	the local filename for one link

internal/article does the fetching and HTML parsing; internal/fetch does the downloading. Because Article is read at run time, --list needs network access too.

WantFiles exists because a partial crawl is otherwise silent: parser will build a short database from whatever files are present, and only the row-count guard would notice, after the fact. A page that changes shape stops the crawl instead.

Local filenames are load-bearing

parser sorts its input files and inserts with INSERT OR REPLACE, which is last-wins, so filenames decide which row survives a duplicate exam number. A re-crawl that names files differently can produce a database with the same row count and different content, which the assembler's row-count guard would not catch.

Each source therefore pins its local names, and crawler/internal/sources/sources_test.go checks every source against its committed data/<id>/ in both directions — every name it would write exists, and every file present is accounted for.

2017 derives its names from the province name in the link text, transliterated to ASCII (the CDN's own names are inconsistent: Angiang.xls, 1BaRiaVungTau.xls, 23HaiPhong.xls). 2016 keeps the server-assigned names verbatim, since it did not choose them.

Source Excel shapes

2017 has one layout:

Col	Name	Content
0	HO_TEN	full name in Vietnamese
1	NGAY_SINH	dd/mm/yyyy
2	SOBAODANH	8-digit string, first 2 digits = province
3	DIEM_THI	concatenated per-subject scores, e.g. "Toán: 6.80 Ngữ văn: 5.25 …"

2016 has four layouts across its 119 files, chosen per sheet at runtime via format_detection: thptqg2016 in its config:

Format	Detected by	Notes
separate-scores	row[0] == "SBD" and row[2] == "TOAN"	one column per subject; scores read directly, no regex
mapped	header contains SOBAODANH/SBD plus DIEM_THI	column indices derived from the header
subject-columns	header names three or more subject columns	university-cluster files; identity and scores both resolved by header name
default	no recognised header	positional 6-column layout

The header may sit below a title block, so the first five rows are searched for it; rows above it are not data.

subject-columns covers two spellings. The ĐH Công nghiệp Thực phẩm file names its columns TO VA LI HO SI SU DI NN with the language code in Môn NN; the three ĐH Cần Thơ files publish cdiem1..cdiem8, numbered in the order the 2016 exam was sat — Toán, Ngoại ngữ, Ngữ văn, Vật lí, Địa lí, Hóa học, Lịch sử, Sinh học — with the language code in ngoaingu. The language score is filed under the subject its N1..N6 code names.

The mapped, default and subject-columns layouts also carry GIOI_TINH (and mapped/default TEN_CUMTHI), which is why only 2016 populates those columns.

Score text parsing

SCORE_PATTERNS in parser/internal/schema/schema.go defines one regex per subject, and all 16 run against every dataset. A subject a given exam year did not offer simply never matches and stays NULL.

Column	Source label	English	2016	2017
toan	Toán	Math	✓	✓
ngu_van	Ngữ văn	Literature	✓	✓
vat_ly	Vật lí	Physics	✓	✓
hoa_hoc	Hóa học	Chemistry	✓	✓
sinh_hoc	Sinh học	Biology	✓	✓
khtn	KHTN	Natural Sciences combined	—	✓
lich_su	Lịch sử	History	✓	✓
dia_ly	Địa lí	Geography	✓	✓
gdcd	GDCD	Civic Education	—	✓
khxh	KHXH	Social Sciences combined	—	✓
tieng_anh	Tiếng Anh	English	✓	✓
tieng_phap	Tiếng Pháp	French	✓	✓
tieng_nga	Tiếng Nga	Russian	✓	✓
tieng_duc	Tiếng Đức	German	✓	✓
tieng_nhat	Tiếng Nhật	Japanese	✓	✓
tieng_trung	Tiếng Trung	Chinese	✓	✓

The foreign-language recovery

Before the parser was unified, the 2016 config listed 12 subject patterns and the 2017 configs listed 14. Neither list was complete: candidates could sit German, Japanese and Russian in both years, so 1,691 students ended up with no foreign-language score at all.

Running all 16 patterns everywhere recovered them: 182 × tieng_nga in 2016, and 93 × tieng_duc plus 512 × tieng_nhat in 2017.

Four things established that these are real scores and not false regex matches, checked when the change was made:

• Every student holds zero or exactly one foreign language, never two, so nothing was double-counted.
• Each affected student had all language columns NULL beforehand — e.g. SBD 01003198 went from all-NULL to tieng_duc = 8.
• The counts tracked dataset size across the three 2017 publications that then existed (93/85/22 German, 512/484/313 Japanese).
• Every recovered value is an ordinary exam score in the 0–10 range.

Row counts and the non-NULL counts of all 18 pre-existing columns were unchanged across every dataset; only the four language columns gained values.

Per-dataset quirks

id	Sheet strategy	Guards
2016	all sheets	per-file format detection; header-token rows rejected
2017	all sheets — Hà Nội and HCM overflow	none

Overflow-sheet gotcha

The old .xls format caps at 65,536 rows per sheet. Hanoi and Ho Chi Minh City exceed that, so Excel splits the overflow into Sheet2. Reading only Sheet1 silently drops 13,720 students (Hanoi +7,275, HCM +6,445). That is what sheet_mode = "all" exists for.

Expected row counts

id	Source rows	Skipped	DB rows
2016	877,460	0	877,460
2017	861,068	0	861,068

Verifying a rebuild

The assembler verifies itself: each database's row count must match the figure in the table above, and each .sqlite3 must be at least 90% of its usual size, or the build fails rather than publishing. That guard is the reason a truncated dataset cannot reach the site with a green pipeline.

For a deeper check, assemble verify compares two sets of built databases field by field — row counts, per-column non-NULL counts, schema metadata, and a full-table SHA-256 over every row ordered by so_bao_danh:

cp -r .build/public/db /tmp/before      # keep the databases you have
go -C assembler run ./cmd/assemble db   # rebuild
go -C assembler run ./cmd/assemble verify /tmp/before .build/public/db

Each side is a directory of <id>.sqlite3; a gzipped database from before the switch to range requests is still expanded to a temporary file automatically. It exits non-zero on any mismatch, names the first differing rows and columns, and fails rather than skipping when a dataset is absent from either side — silently comparing one of two datasets is how a gate passes without proving anything.

This is the only check on database content. The reader oracle below covers reading the spreadsheets and the row-count guard covers how many rows came out, but a change in transform or writer logic can alter what is in those rows while both of those still pass.

parser/internal/reader additionally carries a frozen oracle of per-file cell-dump hashes covering all 182 inputs; go -C parser test ./... fails if any single cell of any input file reads differently.

Refreshing the 2017 data

rm data/2017/*.xls
go -C crawler run ./cmd/crawl 2017
go -C assembler run ./cmd/assemble db 2017

The row-count guard in build:db confirms the rebuild matches the expected total. That guard checks the count only, so if the crawl was expected to change the data, compare content with assemble verify against a copy of the previous database rather than trusting the count.

Removed scripts

check-duplicates.js, diff-datasets.js, db-stats.js and verify-parity.js were dropped with the Rust parser. The first two had been broken since before the repo was unified (a hardcoded Windows path in one, an undeclared better-sqlite3 dependency in the other) and neither had any automated caller. The latter two are superseded by assemble verify, which compares more and cannot silently skip a dataset. differential-parity.mjs, which was that comparator, has itself been folded into the assembler as assembler/internal/verify — the pipeline is Go outside web/, and the port also fixed a weakness: the JavaScript version hashed each row's fields joined bare, so a value shifted across a column boundary produced the same digest.

crawl-baotintuc.js was not dropped but rewritten as the Go crawler/ module, producing the same local filenames. Two changes beyond the port:

• It carried its 63 links as a hardcoded array. The Go version reads them from the article instead, so the list cannot drift from what was published.
• Downloads land on a .part file and are renamed on completion. Writing straight to the destination left a truncated file after an interrupted run, and since the skip check only tests for a non-empty file, every later run would skip it — the corruption was permanent and silent.