askrobots / dbbasic-object-server
Public
Runs live, versioned Python application objects on one VM, with a suite of apps: projects, notes, tasks, files and more.
danq9
signup: the person chooses the password and name on the confirm page
83c7aa6
2 weeks, 5 days ago
History
object_visitors.py · 12.0 KB · 296 lines
| 1 | """Who actually turned up: visitors, bots, and us, counted apart. |
| 2 | |
| 3 | The pure half of visitor analytics -- one pass over page_views rows, |
| 4 | no data directory, no I/O. |
| 5 | |
| 6 | The question this exists to answer is "I talked about the thing today, |
| 7 | did anyone come?", and a raw unique-IP count answers it badly. On this |
| 8 | server 45% of distinct paths were hit exactly once, by vulnerability |
| 9 | scanners probing for /wp-login.php and /index1.php from a fresh address |
| 10 | each time. Counting those as visitors turns a quiet day into a busy one |
| 11 | and makes the number useless precisely when somebody is relying on it. |
| 12 | |
| 13 | So traffic is classified into three kinds and never silently merged: |
| 14 | |
| 15 | **operator** -- our own. The `is_owner` flag already marks owner IPs and |
| 16 | admin-token requests, so deploys and scripts stop inflating the count. |
| 17 | Shown, not hidden: seeing your own hits is how you sanity-check that the |
| 18 | page is recording anything at all. |
| 19 | |
| 20 | **bot** -- a client that never behaved like a person. The test is |
| 21 | behavioural rather than a user-agent blocklist, because the blocklist is |
| 22 | always out of date and the behaviour is not: a scanner walks a list of |
| 23 | URLs that do not exist here and collects 404s. |
| 24 | |
| 25 | A successful page load alone is NOT enough, and that correction came from |
| 26 | real data: `/` returns 200 to everyone, so the first version of this |
| 27 | promoted every prober that touched the homepage into the visitor column. |
| 28 | An address must also have loaded a second distinct page, arrived from a |
| 29 | referrer, or carried a session cookie -- any one of which a person does |
| 30 | and a prober does not. |
| 31 | |
| 32 | A declared crawler stays a bot even when it fetches something real: |
| 33 | Googlebot is honest about what it is, and honesty should not promote it. |
| 34 | So does anything sending an attack payload in a header -- one arrived |
| 35 | here as a Log4Shell probe with an AWS-credential exfiltration template in |
| 36 | the Referer, and was counted as a visitor because it had also loaded the |
| 37 | front page. |
| 38 | |
| 39 | **visitor** -- everything else. A human, probably. |
| 40 | |
| 41 | "Unique" means distinct IP, and that is an approximation worth stating |
| 42 | out loud rather than dressing up: an office behind one NAT is one |
| 43 | visitor, a phone moving between wifi and cellular is two. It stays the |
| 44 | unit HERE even though there is now a first-party visitor cookie |
| 45 | (object_analytics), and that is a deliberate split rather than an |
| 46 | oversight: this fold classifies BEHAVIOUR, and behaviour belongs to an |
| 47 | address -- a scanner does not accept cookies, so classifying by token |
| 48 | would put every bot in one bucket called "no token" and make the |
| 49 | visitor/bot line unanswerable. The cookie answers a different question, |
| 50 | new versus returning, and object_conversions owns that fold. The |
| 51 | alternative to both -- a cross-site identifier, or a fingerprint -- is a |
| 52 | thing this system deliberately does not have. |
| 53 | """ |
| 54 | |
| 55 | from __future__ import annotations |
| 56 | |
| 57 | from collections import defaultdict |
| 58 | from datetime import datetime, timedelta, timezone |
| 59 | from typing import Any, Iterable |
| 60 | |
| 61 | # Clients that say what they are. Matched case-insensitively against a |
| 62 | # substring of the user agent -- deliberately short, because this list is |
| 63 | # a courtesy to honest crawlers, not the mechanism that catches scanners. |
| 64 | # The behavioural test below is what does that work. |
| 65 | _DECLARED_BOTS = ( |
| 66 | "bot", "crawler", "spider", "slurp", "curl/", "wget", "python-requests", |
| 67 | "httpx", "go-http-client", "scrapy", "headlesschrome", "monitoring", |
| 68 | "uptime", "pingdom", "semrush", "ahrefs", "facebookexternalhit", |
| 69 | ) |
| 70 | |
| 71 | # Paths that a person does not "visit" -- fetched by a page, not chosen by |
| 72 | # a human. They still count as traffic; they just cannot be the ONLY thing |
| 73 | # an address did and still make it a visitor. |
| 74 | _NON_PAGE_PREFIXES = ("/api/", "/collections/", "/objects/", "/style", "/nav") |
| 75 | |
| 76 | # Attack payloads seen in a header. Not a defence -- none of these work |
| 77 | # against this server, and the JNDI one is a Java vulnerability probing a |
| 78 | # Python process -- but a client that sends one is definitively not a |
| 79 | # person, and saying so is what keeps it out of the visitor column. The |
| 80 | # real example that prompted this arrived in a Referer header: |
| 81 | # ${jndi:ldap://…}AWS_SECRET_ACCESS_KEY=${env:AWS_SECRET_ACCESS_KEY:-}… |
| 82 | # and was counted as a visitor, because it had also loaded the homepage. |
| 83 | _HOSTILE_MARKERS = ( |
| 84 | "${jndi:", "${env:", "../../", "..%2f", "<script", "union select", |
| 85 | "' or '1'='1", "/etc/passwd", "%00", "{{", "$(", "`", |
| 86 | ) |
| 87 | |
| 88 | |
| 89 | def hostile(*parts: str) -> bool: |
| 90 | """True when any header carries an attack payload.""" |
| 91 | haystack = " ".join(_text(part) for part in parts).lower() |
| 92 | return any(marker in haystack for marker in _HOSTILE_MARKERS) |
| 93 | |
| 94 | |
| 95 | OPERATOR, BOT, VISITOR = "operator", "bot", "visitor" |
| 96 | |
| 97 | |
| 98 | def _text(value: Any) -> str: |
| 99 | return str(value if value is not None else "").strip() |
| 100 | |
| 101 | |
| 102 | def _truthy(value: Any) -> bool: |
| 103 | return _text(value).lower() in ("true", "1", "yes", "on") |
| 104 | |
| 105 | |
| 106 | def _status(value: Any) -> int: |
| 107 | try: |
| 108 | return int(_text(value) or 0) |
| 109 | except (TypeError, ValueError): |
| 110 | return 0 |
| 111 | |
| 112 | |
| 113 | def is_page_path(path: str) -> bool: |
| 114 | """A path a person could have chosen to visit.""" |
| 115 | clean = _text(path) or "/" |
| 116 | return not clean.startswith(_NON_PAGE_PREFIXES) |
| 117 | |
| 118 | |
| 119 | def declared_bot(user_agent: str) -> bool: |
| 120 | lowered = _text(user_agent).lower() |
| 121 | if not lowered: |
| 122 | # No user agent at all is a script, not a browser. Every real |
| 123 | # browser sends one. |
| 124 | return True |
| 125 | return any(token in lowered for token in _DECLARED_BOTS) |
| 126 | |
| 127 | |
| 128 | def classify(rows: Iterable[dict]) -> dict[str, str]: |
| 129 | """Decide what each IP was, from everything it did. |
| 130 | |
| 131 | Classification is per ADDRESS rather than per request, because the |
| 132 | question is "how many of these were people" and a person generates |
| 133 | many rows. |
| 134 | |
| 135 | A successful page load is necessary but NOT sufficient, and that |
| 136 | correction came from the data: `/` returns 200 to everyone, so the |
| 137 | first version promoted every scanner that touched the homepage into |
| 138 | the visitor column. On a real week that was a large share of the |
| 139 | "direct traffic" figure. So an address also has to have done one |
| 140 | thing a prober does not: |
| 141 | |
| 142 | * loaded MORE THAN ONE distinct page, or |
| 143 | * arrived from somewhere (a referrer), or |
| 144 | * carried a session cookie. |
| 145 | |
| 146 | A person browsing does at least one of those; something hitting root |
| 147 | and leaving does none. The rule is deliberately generous -- one real |
| 148 | page plus any one signal -- because over-counting a shy visitor is a |
| 149 | smaller error than reporting a busy day that never happened. |
| 150 | """ |
| 151 | seen_pages: dict[str, set[str]] = defaultdict(set) |
| 152 | signalled: set[str] = set() |
| 153 | declared: set[str] = set() |
| 154 | operator: set[str] = set() |
| 155 | everyone: set[str] = set() |
| 156 | |
| 157 | for row in rows: |
| 158 | ip = _text(row.get("ip")) |
| 159 | if not ip: |
| 160 | continue |
| 161 | everyone.add(ip) |
| 162 | if _truthy(row.get("is_owner")): |
| 163 | operator.add(ip) |
| 164 | referrer = _text(row.get("referrer")) |
| 165 | if (declared_bot(row.get("user_agent")) |
| 166 | or hostile(referrer, row.get("user_agent"), row.get("path"))): |
| 167 | declared.add(ip) |
| 168 | if (200 <= _status(row.get("status")) < 400 |
| 169 | and is_page_path(row.get("path"))): |
| 170 | seen_pages[ip].add(_text(row.get("path")) or "/") |
| 171 | if referrer or _text(row.get("session_id")): |
| 172 | signalled.add(ip) |
| 173 | |
| 174 | kinds: dict[str, str] = {} |
| 175 | for ip in everyone: |
| 176 | pages = seen_pages.get(ip, set()) |
| 177 | looks_human = bool(pages) and (len(pages) > 1 or ip in signalled) |
| 178 | if ip in operator: |
| 179 | kinds[ip] = OPERATOR |
| 180 | elif ip in declared or not looks_human: |
| 181 | kinds[ip] = BOT |
| 182 | else: |
| 183 | kinds[ip] = VISITOR |
| 184 | return kinds |
| 185 | |
| 186 | |
| 187 | def _bucket(stamp: str, *, hourly: bool) -> str: |
| 188 | text = _text(stamp) |
| 189 | if len(text) < 10: |
| 190 | return "" |
| 191 | return text[:13].replace("T", " ") if hourly else text[:10] |
| 192 | |
| 193 | |
| 194 | def summarize( |
| 195 | rows: Iterable[dict], |
| 196 | *, |
| 197 | now: datetime | None = None, |
| 198 | days: int = 7, |
| 199 | host: str = "", |
| 200 | ) -> dict[str, Any]: |
| 201 | """Visitors, views and bots -- hourly for today, daily for the window. |
| 202 | |
| 203 | Returns {"today", "days", "hours", "totals", "referrers", "hosts", ...}. |
| 204 | Counted in ONE pass with sets per bucket: the alternative, a distinct |
| 205 | count per bucket computed by re-scanning, is the shape that made a |
| 206 | rollup rewrite 1818 rows every five minutes on this very server. |
| 207 | |
| 208 | `host` narrows to one site. One process here answers for a marketing |
| 209 | domain and the app domain both, and a combined figure answers neither |
| 210 | "did anyone read the pitch" nor "did anyone try the product". The |
| 211 | per-host split is always returned even when not filtering, because |
| 212 | the first useful thing to know is which site they went to. |
| 213 | """ |
| 214 | rows = list(rows) |
| 215 | if host: |
| 216 | rows = [row for row in rows if _text(row.get("host")) == host] |
| 217 | kinds = classify(rows) |
| 218 | now = now or datetime.now(timezone.utc) |
| 219 | today = now.strftime("%Y-%m-%d") |
| 220 | earliest = (now - timedelta(days=days - 1)).strftime("%Y-%m-%d") |
| 221 | |
| 222 | views: dict[tuple[str, str], int] = defaultdict(int) |
| 223 | uniques: dict[tuple[str, str], set[str]] = defaultdict(set) |
| 224 | hour_views: dict[tuple[str, str], int] = defaultdict(int) |
| 225 | hour_uniques: dict[tuple[str, str], set[str]] = defaultdict(set) |
| 226 | referrers: dict[str, set[str]] = defaultdict(set) |
| 227 | landing: dict[str, set[str]] = defaultdict(set) |
| 228 | hosts: dict[str, set[str]] = defaultdict(set) |
| 229 | |
| 230 | for row in rows: |
| 231 | ip = _text(row.get("ip")) |
| 232 | stamp = _text(row.get("created_at")) |
| 233 | day = _bucket(stamp, hourly=False) |
| 234 | if not ip or not day or day < earliest: |
| 235 | continue |
| 236 | kind = kinds.get(ip, VISITOR) |
| 237 | |
| 238 | views[(day, kind)] += 1 |
| 239 | uniques[(day, kind)].add(ip) |
| 240 | |
| 241 | if day == today: |
| 242 | hour = _bucket(stamp, hourly=True) |
| 243 | hour_views[(hour, kind)] += 1 |
| 244 | hour_uniques[(hour, kind)].add(ip) |
| 245 | |
| 246 | if kind != VISITOR: |
| 247 | continue |
| 248 | hosts[_text(row.get("host")) or "(not recorded)"].add(ip) |
| 249 | referrer = _text(row.get("referrer")) |
| 250 | # An empty referrer is "typed it, or came from a private link" -- |
| 251 | # worth showing as its own row rather than dropped, because for a |
| 252 | # server nobody has linked to yet it is the whole story. |
| 253 | referrers[referrer or "(direct or private link)"].add(ip) |
| 254 | if is_page_path(row.get("path")) and 200 <= _status(row.get("status")) < 400: |
| 255 | landing[_text(row.get("path")) or "/"].add(ip) |
| 256 | |
| 257 | def _series(bucket_views, bucket_uniques, keys): |
| 258 | return [{ |
| 259 | "bucket": key, |
| 260 | "visitors": len(bucket_uniques.get((key, VISITOR), ())), |
| 261 | "views": bucket_views.get((key, VISITOR), 0), |
| 262 | "bots": len(bucket_uniques.get((key, BOT), ())), |
| 263 | "bot_views": bucket_views.get((key, BOT), 0), |
| 264 | "operator_views": bucket_views.get((key, OPERATOR), 0), |
| 265 | } for key in keys] |
| 266 | |
| 267 | day_keys = [(now - timedelta(days=offset)).strftime("%Y-%m-%d") |
| 268 | for offset in range(days - 1, -1, -1)] |
| 269 | hour_keys = [f"{today} {hour:02d}" for hour in range(24)] |
| 270 | |
| 271 | totals: dict[str, Any] = {} |
| 272 | for kind in (VISITOR, BOT, OPERATOR): |
| 273 | ips = {ip for ip, k in kinds.items() if k == kind} |
| 274 | totals[kind] = { |
| 275 | "unique": len(ips), |
| 276 | "views": sum(count for (day, k), count in views.items() |
| 277 | if k == kind and day >= earliest), |
| 278 | } |
| 279 | |
| 280 | return { |
| 281 | "today": today, |
| 282 | "days": _series(views, uniques, day_keys), |
| 283 | "hours": _series(hour_views, hour_uniques, hour_keys), |
| 284 | "totals": totals, |
| 285 | "referrers": sorted( |
| 286 | ({"referrer": ref, "visitors": len(ips)} for ref, ips in referrers.items()), |
| 287 | key=lambda row: (-row["visitors"], row["referrer"]))[:12], |
| 288 | "hosts": sorted( |
| 289 | ({"host": name, "visitors": len(ips)} for name, ips in hosts.items()), |
| 290 | key=lambda row: (-row["visitors"], row["host"])), |
| 291 | "host": host, |
| 292 | "landing": sorted( |
| 293 | ({"path": path, "visitors": len(ips)} for path, ips in landing.items()), |
| 294 | key=lambda row: (-row["visitors"], row["path"]))[:12], |
| 295 | "window_days": days, |
| 296 | } |