listRepos is too slow

[snarfed]

I'm seeing it take 20+s in prod right now for limit=1000, which is way too slow.

$ time curl -v 'https://atproto.brid.gy/xrpc/com.atproto.sync.listRepos?limit=1000'
...
0.022u 0.028s 0:22.79 0.1%	0+0k 0+0io 2pf+0w

Not sure why though. All it does is call storage.load_repos and then construct the output, and all load_repos does is query AtpRepos, load their head AtpBlocks, and then return.

arroba/arroba/xrpc_sync.py

Lines 74 to 94 in 75cefb6

	def list_repos(input, limit=500, cursor=None):
	"""Handler for ``com.atproto.sync.listRepos`` XRPC method."""
	STATUSES = {'tombstoned': 'deactivated'}
	repos = []
	for repo in server.storage.load_repos(limit=limit, after=cursor, minimal=True):
	repo_obj = {
	'did': repo.did,
	'head': repo.head.cid.encode('base32'),
	'rev': util.int_to_tid(repo.head.seq, clock_id=0),
	'active': repo.status is None,
	}
	if repo.status:
	repo_obj['status'] = STATUSES.get(repo.status) or repo.status
	repos.append(repo_obj)
	ret = {'repos': repos}
	if len(repos) == limit:
	ret['cursor'] = repos[-1]['did']
	return ret

arroba/arroba/datastore_storage.py

Lines 718 to 740 in 75cefb6

	def load_repos(self, after=None, limit=500, minimal=False):
	query = AtpRepo.query()
	if after:
	query = query.filter(AtpRepo.key > AtpRepo(id=after).key)
	# duplicates parts of Repo.load but batches reading blocks from storage
	atp_repos = query.fetch(limit=limit)
	cids = [CID.decode(r.head) for r in atp_repos]
	blocks = self.read_many(cids) # dict mapping CID to block
	heads = [blocks[cid] for cid in cids]
	if minimal:
	return [Repo(storage=self, head=head, status=atp_repo.status)
	for atp_repo, head in zip(atp_repos, heads)]
	# MST.load doesn't read from storage
	msts = [MST.load(storage=self, cid=block.decoded['data']) for block in heads]
	return [Repo(storage=self, mst=mst, head=head, status=atp_repo.status,
	handle=atp_repo.handles[0] if atp_repo.handles else None,
	signing_key=atp_repo.signing_key,
	rotation_key=atp_repo.rotation_key)
	for atp_repo, head, mst in zip(atp_repos, heads, msts)]

Note the minimal=True. I originally though the private key parsing and loading was the culprit, so I now skip that, but that didn't really help. 😕

[snarfed]

cc @verdverm

[verdverm]

In the official PDS, this is a query to the account database, not requiring opening every account repo.

IIRC, they keep the min info updated in that account sqlite when new ops come in

https://github.com/bluesky-social/atproto/tree/main/packages/pds/src/account-manager

[snarfed]

Hah, sure! Bridgy Fed is designed differently, repo metadata and blocks are stored in the same datastore, so there's no separate operation to "open" a repo.

[snarfed]

It's better now, 12-14s warm, but still not great.