Fan-out on write.

# One pipeline per shard; followers are
# already grouped by shard.
pipe = shard.pipeline(transaction=False)
for uid in followers_on_this_shard:
    key = f"feed:{uid}"
    # LPUSHX pushes only if the feed exists:
    # never create a stub feed.
    pipe.lpushx(key, f"{post_id}:{author_id}")
    pipe.ltrim(key, 0, 499)  # newest 500
pipe.execute()
# LTRIM right after LPUSH removes one
# element: O(1) in practice. But a
# redelivered batch pushes the post twice.

# Plain ZADD on a missing key creates a
# one-entry feed the read path would trust.
# So append only if the key exists.
APPEND = shard.register_script("""
  if redis.call('EXISTS', KEYS[1]) == 0
    then return 0 end
  -- same member twice = one entry
  redis.call('ZADD', KEYS[1],
             ARGV[1], ARGV[2])
  -- keep the newest 500
  redis.call('ZREMRANGEBYRANK',
             KEYS[1], 0, -501)
  return 1""")

# The ID's 41-bit ms timestamp + epoch.
score = created_ms(post_id)
member = f"{post_id}:{author_id}"
pipe = shard.pipeline(transaction=False)
for uid in followers_on_this_shard:
    APPEND(keys=[f"feed:{uid}"],
           args=[score, member], client=pipe)
pipe.execute()
# Scores are doubles, exact only to 2^53:
# a full 64-bit post_id would lose its low
# bits, so score by the timestamp inside it.