Fold per-place copies of one host into one role that lists every place with its own link
The reader shows every place, each linked to its copy, and the teaser marked as a short description when there is no text. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
1 parent
3adafd8950
commit
3d1fed7c0e
6 files changed
+336
-16
No files matched your search
+97
-1
@@ -17,14 +17,30 @@ rows that each match all the others, never a chain: a row with no place
|
||||
matches Belgium and the Netherlands, and must not make them one role.
|
||||
|
||||
Each group keeps one copy (`keep_order`, smallest wins) and the rest point at
|
||||
it. Pure by design: no store, no SQL, no network, so the rule is tested on
|
||||
it.
|
||||
|
||||
A second rule folds what the first leaves apart on purpose: one posting per
|
||||
place. A host that lists one job once per office, each with its own id, gives
|
||||
rows with the same company and title whose towns never meet. `fold` merges
|
||||
those rows of one host into the kept role of the first group, when the texts
|
||||
do not say they are different jobs (`similar`), and reports them in
|
||||
`Folded.placed`: the store gives the kept role each such copy's place with the
|
||||
copy's own link. This holds for every host, not only readers that declare
|
||||
`READS.places == "copies"`: on the live database 1,739 company-and-title
|
||||
groups (5,024 roles) are this, in 18 hosts, 1,700 of them within one host, and
|
||||
Greenhouse, Ashby and SmartRecruiters carry most, not HiBob. Rows whose texts
|
||||
share fewer than `SAME_TEXT` of their words, or without a link to keep, or
|
||||
listed by two hosts, stay apart. Pure by design: no store, no SQL, no network, so the rule is tested on
|
||||
the strings boards really write and the storage layer only feeds it rows.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from collections.abc import Iterable
|
||||
from dataclasses import dataclass
|
||||
from functools import lru_cache
|
||||
from typing import NamedTuple
|
||||
|
||||
from app import places
|
||||
from app.models import country_of
|
||||
@@ -48,6 +64,11 @@ class Copy:
|
||||
host: str
|
||||
has_description: bool
|
||||
posted_at: str | None
|
||||
#: The start of the description, else the teaser, for `similar`; '' when
|
||||
#: the row has neither.
|
||||
text: str = ""
|
||||
#: The link that opens this row on its host.
|
||||
url: str = ""
|
||||
|
||||
|
||||
def _country(c: Copy) -> str:
|
||||
@@ -118,3 +139,78 @@ def assign(rows: Iterable[Copy]) -> dict[str, str | None]:
|
||||
if row is not keeper:
|
||||
result[row.uid] = keeper.uid
|
||||
return result
|
||||
|
||||
|
||||
#: The share of words two texts must have in common (Jaccard, over the first
|
||||
#: 3,000 characters) for one job in two places. Measured on the live
|
||||
#: database: copies of a posting share nearly all of it, two jobs that
|
||||
#: happen to share a title share little more than boilerplate, and a
|
||||
#: translation shares none, which keeps it apart.
|
||||
SAME_TEXT = 0.3
|
||||
|
||||
TEXT_CHARS = 3000
|
||||
|
||||
|
||||
@lru_cache(maxsize=8192)
|
||||
def _words(text: str) -> frozenset[str]:
|
||||
return frozenset(re.findall(r"\w{3,}", text[:TEXT_CHARS].lower()))
|
||||
|
||||
|
||||
def similar(a: str, b: str) -> bool:
|
||||
"""Whether two texts can be one job: a text that is missing says nothing
|
||||
against it, else they share at least `SAME_TEXT` of their words."""
|
||||
wa, wb = _words(a), _words(b)
|
||||
if not wa or not wb:
|
||||
return True
|
||||
return len(wa & wb) / len(wa | wb) >= SAME_TEXT
|
||||
|
||||
|
||||
class Folded(NamedTuple):
|
||||
"""`assign`'s answer plus the rows that are one place of another role:
|
||||
`placed` holds the uid of each copy that `fold` merged by place."""
|
||||
|
||||
dup_of: dict[str, str | None]
|
||||
placed: frozenset[str]
|
||||
|
||||
|
||||
def fold(rows: Iterable[Copy]) -> Folded:
|
||||
"""`assign`, then the per-place copies of each host folded in.
|
||||
|
||||
Within a `(company_key, title_key)` the kept rows of `assign` are grouped
|
||||
again, in `keep_order`: a row joins the first group whose members are all
|
||||
of its host and `similar` to it. Only a row with a link can be folded, as
|
||||
its link is what the kept role keeps. Every row of a merged group points
|
||||
at its first row, the rows that pointed at a folded one included."""
|
||||
rows = list(rows)
|
||||
result = assign(rows)
|
||||
by_uid = {r.uid: r for r in rows}
|
||||
parts: dict[tuple[str, str], list[Copy]] = {}
|
||||
for uid, dup in result.items():
|
||||
row = by_uid[uid]
|
||||
if dup is None and row.company_key and row.url:
|
||||
parts.setdefault((row.company_key, row.title_key), []).append(row)
|
||||
moved: dict[str, str] = {}
|
||||
for part in parts.values():
|
||||
groups: list[list[Copy]] = []
|
||||
for row in sorted(part, key=keep_order):
|
||||
for group in groups:
|
||||
if all(m.host == row.host and similar(m.text, row.text) for m in group):
|
||||
group.append(row)
|
||||
break
|
||||
else:
|
||||
groups.append([row])
|
||||
for keeper, *others in groups:
|
||||
moved.update((other.uid, keeper.uid) for other in others)
|
||||
for uid, dup in result.items():
|
||||
if uid in moved:
|
||||
result[uid] = moved[uid]
|
||||
elif dup in moved:
|
||||
result[uid] = moved[dup]
|
||||
return Folded(result, frozenset(moved))
|
||||
|
||||
|
||||
def place_line(texts: Iterable[str]) -> str:
|
||||
"""The line a feed row shows for a role's places: the first, and "+N" for
|
||||
the N others. Repeated texts count once."""
|
||||
found = list(dict.fromkeys(t.strip() for t in texts if t and t.strip()))
|
||||
return f"{found[0]} +{len(found) - 1}" if len(found) > 1 else "".join(found)
|
||||
@@ -942,6 +942,7 @@ body[data-reader="open"] .company-strip { display: none; }
|
||||
line-height: 1.4;
|
||||
}
|
||||
.also a b { display: block; color: var(--faint); font-size: .78rem; font-weight: 600; }
|
||||
.also span { display: block; margin-bottom: .55rem; color: var(--dim); font-size: .92rem; line-height: 1.4; }
|
||||
@media (hover: hover) {
|
||||
.also a:hover { color: var(--fg); }
|
||||
}
|
||||
|
||||
+42
-2
@@ -885,6 +885,42 @@
|
||||
</div>`;
|
||||
}
|
||||
|
||||
/* The role's places in the host's order, one entry for each place text a
|
||||
link opens: the copies' own links, and the role's link for its own. */
|
||||
function placeEntries(r) {
|
||||
const seen = new Set();
|
||||
const own = safeUrl(r.original_url);
|
||||
const out = [];
|
||||
for (const p of Array.isArray(r.places) ? r.places : []) {
|
||||
const url = p.url ? safeUrl(p.url) : own;
|
||||
const key = `${p.url || ""}\u0001${p.position}\u0001${p.text}`;
|
||||
if (!p.text || seen.has(key)) continue;
|
||||
seen.add(key);
|
||||
out.push({ text: p.text, url });
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
/* The place line of the reader: the first place and "+N" for the others,
|
||||
as the feed's row shows it. */
|
||||
function placeLine(r) {
|
||||
const texts = [...new Set(placeEntries(r).map((e) => e.text))];
|
||||
return texts.length > 1 ? `${texts[0]} +${texts.length - 1}` : r.place || texts[0] || "";
|
||||
}
|
||||
|
||||
/* Every place of a role that has several, each a link to its own posting,
|
||||
laid out as "Also posted". */
|
||||
function placesList(r) {
|
||||
const entries = placeEntries(r);
|
||||
if (new Set(entries.map((e) => e.text)).size < 2) return "";
|
||||
return `<div class="also places">
|
||||
<h2 class="reader-label">Places</h2>
|
||||
${entries.map((e) => e.url
|
||||
? `<a href="${esc(e.url)}" target="_blank" rel="noopener">${esc(e.text)} ↗</a>`
|
||||
: `<span>${esc(e.text)}</span>`).join("")}
|
||||
</div>`;
|
||||
}
|
||||
|
||||
function drawRow(el, role) {
|
||||
const scored = role.score !== null && role.score !== undefined;
|
||||
const pct = scored ? Math.round(Math.min(1, Math.max(0, role.score)) * 100) : 0;
|
||||
@@ -1243,7 +1279,10 @@
|
||||
: !loaded
|
||||
? `<p class="reader-note">Loading…</p>`
|
||||
: text ? paragraphs(text)
|
||||
: `<p class="reader-note">The board gave no description for this one.</p>`;
|
||||
: (r.teaser || "").trim()
|
||||
? `<p class="reader-note reader-teaser">A short description from the board's
|
||||
list; the board gave no full text.</p>${paragraphs(r.teaser.trim())}`
|
||||
: `<p class="reader-note">The board gave no description for this one.</p>`;
|
||||
const pay = payDisplay(r);
|
||||
const careers = safeUrl(companies.get(r.company_key)?.careers_url);
|
||||
const waiting = state.summarising === state.readingUid;
|
||||
@@ -1264,7 +1303,7 @@
|
||||
${careers ? `<a class="reader-careers" href="${esc(careers)}" target="_blank" rel="noopener"
|
||||
title="The company's careers page">Careers ↗</a>` : ""}
|
||||
<span class="affinity-slot"></span>
|
||||
${r.place ? `<span>${esc(r.place)}</span>` : ""}
|
||||
${placeLine(r) ? `<span>${esc(placeLine(r))}</span>` : ""}
|
||||
${r.first_seen ? `<span title="First seen">${ago(r.first_seen)}</span>` : ""}
|
||||
${badges(r)}
|
||||
${tagChips(r.tags)}
|
||||
@@ -1287,6 +1326,7 @@
|
||||
<div class="prose">${body}</div>
|
||||
${out ? `<p class="reader-out"><a href="${esc(out)}" target="_blank" rel="noopener"
|
||||
>Read it on ${esc(host(out) || "the board")} ↗</a></p>` : ""}
|
||||
${loaded ? placesList(r) : ""}
|
||||
${loaded ? alsoPosted(r) : ""}
|
||||
</div>
|
||||
<div class="reader-foot">
|
||||
|
||||
+75
-7
@@ -10,7 +10,7 @@ from __future__ import annotations
|
||||
|
||||
import json
|
||||
import sqlite3
|
||||
from collections.abc import Callable, Iterator, MutableMapping, Sequence
|
||||
from collections.abc import Callable, Iterable, Iterator, MutableMapping, Sequence
|
||||
from contextlib import closing, contextmanager
|
||||
from datetime import UTC, datetime, timedelta
|
||||
from pathlib import Path
|
||||
@@ -281,6 +281,13 @@ def _int_or_none(value: Any) -> int | None:
|
||||
_HOST = "COALESCE(json_extract(raw, '$.provider'), json_extract(raw, '$.board'), '')"
|
||||
|
||||
|
||||
#: A copy that is not one place of its kept role: the copies a role lists
|
||||
#: under "Also posted" and counts as "x N". `{kept}` is the kept role's uid
|
||||
#: as SQL, and a place copy is the one whose link the kept role holds.
|
||||
_NOT_A_PLACE = ("original_url NOT IN (SELECT url FROM role_places "
|
||||
"WHERE uid = {kept} AND url != '')")
|
||||
|
||||
|
||||
class DupeCounts(NamedTuple):
|
||||
"""What `rebuild_dupes` left and did: kept roles with at least one copy,
|
||||
open roles that are a copy, and rows whose `dup_of` it rewrote."""
|
||||
@@ -290,8 +297,35 @@ class DupeCounts(NamedTuple):
|
||||
changed: int
|
||||
|
||||
|
||||
def _assign_only(rows: Iterable[dupes.Copy]) -> dupes.Folded:
|
||||
return dupes.Folded(dupes.assign(rows), frozenset())
|
||||
|
||||
|
||||
def _sync_copy_places(conn: sqlite3.Connection, placed: dict[str, str | None]) -> None:
|
||||
"""Make the `role_places` rows with a `url` those of the copies `placed`
|
||||
(`copy uid -> kept uid`): each own place of a copy, on the kept role,
|
||||
with the copy's link. Written only when they differ from what is stored,
|
||||
so an unchanged database is not touched. A copy that closed, or a group
|
||||
that dissolved, is no longer in `placed`, and its rows go."""
|
||||
want: set[tuple[str, int, str, str, int | None, str]] = set()
|
||||
if placed:
|
||||
marks = ",".join("?" * len(placed))
|
||||
for uid, pos, text, fold, place_id, url in conn.execute(
|
||||
f"""SELECT rp.uid, rp.position, rp.text, rp.fold, rp.place_id, r.original_url
|
||||
FROM role_places rp JOIN roles r ON r.uid = rp.uid
|
||||
WHERE rp.url = '' AND rp.uid IN ({marks})""", list(placed)):
|
||||
want.add((str(placed[uid]), pos, text, fold, place_id, url))
|
||||
have = set(conn.execute(
|
||||
"SELECT uid, position, text, fold, place_id, url FROM role_places WHERE url != ''"))
|
||||
if want != have:
|
||||
conn.execute("DELETE FROM role_places WHERE url != ''")
|
||||
conn.executemany(
|
||||
"INSERT INTO role_places (uid, position, text, fold, place_id, url) "
|
||||
"VALUES (?, ?, ?, ?, ?, ?)", sorted(want, key=lambda r: (r[0], r[5], r[1], r[2])))
|
||||
|
||||
|
||||
def rebuild_dupes(conn: sqlite3.Connection) -> DupeCounts:
|
||||
"""Recompute `roles.dup_of` from scratch with `app.dupes.assign`, inside
|
||||
"""Recompute `roles.dup_of` from scratch with `app.dupes.fold`, inside
|
||||
the caller's transaction: `Store.rebuild_dupes` and migration 18.
|
||||
|
||||
Only the open roles of a `(company_key, title_key)` with more than one
|
||||
@@ -309,26 +343,40 @@ def rebuild_dupes(conn: sqlite3.Connection) -> DupeCounts:
|
||||
them moves to it unless that agent already voted on it; a vote that
|
||||
cannot move stays where it is. An unchanged group is not touched, so
|
||||
marking a kept role unread again sticks. Applications keep their
|
||||
`role_uid`, and scores of a copy go at the agent's next `score`."""
|
||||
`role_uid`, and scores of a copy go at the agent's next `score`.
|
||||
|
||||
A copy the host published for another place (`dupes.Folded.placed`)
|
||||
leaves its places on the kept role as `role_places` rows with the copy's
|
||||
link, see `_sync_copy_places`."""
|
||||
# Migration 18 runs this on a file that has no `teaser` and no
|
||||
# `role_places` yet: there it groups as before and folds no places.
|
||||
current = any(r[1] == "teaser" for r in conn.execute("PRAGMA table_info(roles)"))
|
||||
text = "COALESCE(NULLIF(description, ''), teaser)" if current else "description"
|
||||
rows = conn.execute(
|
||||
f"""SELECT uid, company_key, title_key, country, place, {_HOST} AS host,
|
||||
description != '' AS described, posted_at
|
||||
description != '' AS described, posted_at,
|
||||
SUBSTR({text}, 1, {dupes.TEXT_CHARS}), original_url
|
||||
FROM roles
|
||||
WHERE closed_at IS NULL AND company_key != ''
|
||||
AND (company_key, title_key) IN (
|
||||
SELECT company_key, title_key FROM roles
|
||||
WHERE closed_at IS NULL AND company_key != ''
|
||||
GROUP BY company_key, title_key HAVING COUNT(*) > 1)""").fetchall()
|
||||
new = dupes.assign(
|
||||
made = dupes.fold if current else _assign_only
|
||||
folded = made(
|
||||
dupes.Copy(uid=r[0], company_key=r[1], title_key=r[2], country=r[3], place=r[4],
|
||||
host=r[5], has_description=bool(r[6]), posted_at=r[7] or None)
|
||||
host=r[5], has_description=bool(r[6]), posted_at=r[7] or None,
|
||||
text=r[8] or "", url=r[9] or "")
|
||||
for r in rows)
|
||||
new = folded.dup_of
|
||||
old: dict[str, str | None] = dict.fromkeys(new)
|
||||
old.update((r[0], r[1]) for r in conn.execute(
|
||||
"SELECT uid, dup_of FROM roles WHERE dup_of IS NOT NULL"))
|
||||
updates = [(new.get(uid), uid) for uid, was in old.items() if new.get(uid) != was]
|
||||
conn.executemany("UPDATE roles SET dup_of = ? WHERE uid = ?", updates)
|
||||
|
||||
if current:
|
||||
_sync_copy_places(conn, {uid: new[uid] for uid in folded.placed})
|
||||
changed = {uid for _, uid in updates}
|
||||
members: dict[str, set[str]] = {}
|
||||
for uid, dup in new.items():
|
||||
@@ -707,7 +755,8 @@ class Store:
|
||||
or a role without any. See `Role.copies`."""
|
||||
return [RoleCopy.from_row(r) for r in self._rows(
|
||||
f"""SELECT uid, {_HOST} AS host, place, original_url, posted_at FROM roles
|
||||
WHERE dup_of = ? AND closed_at IS NULL ORDER BY posted_at, uid""", (uid,))]
|
||||
WHERE dup_of = ? AND closed_at IS NULL AND {_NOT_A_PLACE.format(kept="?")}
|
||||
ORDER BY posted_at, uid""", (uid, uid))]
|
||||
|
||||
def _copy_counts(self, uids: Sequence[str]) -> dict[str, int]:
|
||||
"""`uid -> open copies` for a page of kept roles, one grouped query
|
||||
@@ -720,8 +769,25 @@ class Store:
|
||||
return {str(r["dup_of"]): int(r["n"]) for r in self._rows(
|
||||
"SELECT dup_of, COUNT(*) AS n FROM roles INDEXED BY idx_roles_dup_of "
|
||||
"WHERE dup_of IN (SELECT value FROM json_each(?)) AND closed_at IS NULL "
|
||||
f"AND {_NOT_A_PLACE.format(kept='dup_of')} "
|
||||
"GROUP BY dup_of", (json.dumps(list(uids)),))}
|
||||
|
||||
def _place_lines(self, uids: Sequence[str]) -> dict[str, str]:
|
||||
"""`uid -> the line the feed shows` for a page of roles that have
|
||||
more than one place (`dupes.place_line`); a role with one place or
|
||||
none is absent and keeps `roles.place`."""
|
||||
if not uids:
|
||||
return {}
|
||||
found: dict[str, list[str]] = {}
|
||||
for r in self._rows(
|
||||
"""SELECT uid, text FROM role_places
|
||||
WHERE uid IN (SELECT value FROM json_each(?))
|
||||
ORDER BY uid, url != '', url, position, rowid""",
|
||||
(json.dumps(list(uids)),)):
|
||||
found.setdefault(str(r["uid"]), []).append(str(r["text"]))
|
||||
return {uid: dupes.place_line(texts) for uid, texts in found.items()
|
||||
if len(set(texts)) > 1}
|
||||
|
||||
def rebuild_dupes(self) -> DupeCounts:
|
||||
"""Recompute which open roles are copies of one posting, in one
|
||||
transaction; see the module's `rebuild_dupes`. `BEGIN IMMEDIATE`
|
||||
@@ -966,9 +1032,11 @@ class Store:
|
||||
found.append(RoleRow.from_row(row))
|
||||
by_uid = self._tags_by_uid([f.uid for f in found])
|
||||
copies = self._copy_counts([f.uid for f in found])
|
||||
lines = self._place_lines([f.uid for f in found])
|
||||
for f in found:
|
||||
f.tags = by_uid.get(f.uid, [])
|
||||
f.copies = copies.get(f.uid, 0)
|
||||
f.place = lines.get(f.uid, f.place)
|
||||
return found
|
||||
|
||||
def count_roles(self, company: str | None = None, place: str | None = None,
|
||||
|
||||
Reference in new issue
Block a user