Parsing YouTube's abbreviated counts

One platform, two conventions

This is the thing to settle before you design a YouTube schema, because getting it wrong produces a pipeline that works on one actor and silently corrupts another.

Actor	views / likes	duration	timestamps
Comments	strings, abbreviated	n/a	relative phrase
Channel Videos	integers	seconds, integer	n/a in the sample
Playlist	integers	seconds, integer	n/a in the sample

The comments actor hands back what YouTube painted on the screen. The other two hand back data. Both are documented, neither is wrong, and a team using both will write one parser and apply it twice unless somebody says this out loud.

Abbreviation is lossy and permanent

“3.4K” is not 3,400. It is somewhere between roughly 3,350 and 3,449, and the information that would let you narrow it was discarded before you received the string.

That has a practical consequence: the parsed number must never be the only copy you keep. If you store 3,400 and throw away “3.4K”, you have converted a known approximation into a false precision, and six months later nobody will remember which columns were exact.

import re

def parse_count(s):
    """'3.4K' -> 3400. Approximate by construction."""
    s = (s or "0").strip()
    mult = {"K": 1_000, "M": 1_000_000, "B": 1_000_000_000}.get(s[-1:].upper(), 1)
    return int(float(re.sub(r"[^\d.]", "", s) or 0) * mult)

row = {
    **comment,
    "likes_n": parse_count(comment["likeCount"]),  # derived, approximate
    "likes_raw": comment["likeCount"],             # original, never overwritten
}

Two columns. One you sort on, one you can defend.

The timestamp problem is worse

publishedTime is a relative phrase: “5 months ago”. It is measured from the moment of the scrape, which means the same comment reads differently every time you run it, and two runs cannot be merged on it at all.

There is no absolute timestamp in the record. So the run’s own time has to come from you:

scraped_at = datetime.now(timezone.utc)
rows = [{**c, "scraped_at": scraped_at.isoformat()} for c in items]

Without that, a comment dataset has no chronology. With it, “5 months ago” resolves to a real date and stays resolved.

Where to be strict

Use the parsed numbers for ranking, filtering and thresholds. That is what they are good for and the error does not matter when you are asking which comments are the most liked.

Do not use them for totals you report to somebody. Summing four hundred abbreviated counts compounds four hundred roundings, and the answer looks authoritative in a way it has not earned.

When a figure has to be exact, the video’s own engagement from the Channel Videos actor is an integer, and that is the number to quote.