etag errors in ceph backed instance

[rklra]

Hi,

I was trying to get the package setup using ceph object storage (theoretically s3 compatible) as the backing store but the etag system used to protect against concurrent writes (

sneller/db/build.go

Line 67 in 86e9f11

func getInfo(dst OutputFS, fp string, out blockfmt.Uploader) (string, time.Time, error) {

) does not seem to function with cephs etag system (with the value of e.ETag() being blank) as seen here:

sync: etag "[etag here]" from Stat disagrees with etag

rebuilding sdb with the if statement here:

if e, ok := out.(etagger); ok && e.ETag() != etag {
		return "", time.Time{}, fmt.Errorf("etag %s from Stat disagrees with etag %s", etag, e.ETag())
	}

commented out, queries appear to function normally and syncing continues to work (as long as no concurrent reading/writing to the table is occuring)

full outputs here:
(unmodified version of sdb)

root@avx1 ~/sneller # go run ./cmd/sdb/ -- sync main table
sync: etag "[full etag here]" from Stat disagrees with etag
exit status 1

(modified version of sdb)

root@avx1 ~/sneller # go run ./cmd/sdb/ -- sync main table

(edited version -- worked with no errors)

note on concurrent read/write

in my (limited) testing with concurrent reads/writes occurring while the sync task is running, few to no issues occur with querying data that was synced during a write on my modified instance

(sync during write)

root@avx1 ~/sneller # go run ./cmd/sdb/ -- sync main table
root@avx1 ~/sneller # sdb query -fmt json "SELECT COUNT(*) FROM main.table"
{"count": 76}

(sync after write)

root@avx1 ~/sneller # go run ./cmd/sdb/ -- sync main table
root@avx1 ~/sneller # sdb -v  query -fmt json "SELECT COUNT(*) FROM main.table"
{"count": 87}

(full rebuild of indexes and zion files)

root@avx1 ~/sneller # go run ./cmd/sdb/ -- sync main table
root@avx1 ~/sneller # sdb -v  query -fmt json "SELECT COUNT(*) FROM main.table"
{"count": 87}

[ramondeklein]

I checked the ceph documentation and it looks like the Get Object method doesn't return the ETag header, which differs from the normal S3 API. It looks like ceph does know about the ETag, because the GET method does support the if-match request header. I would file a PR with ceph to return the ETag when listing buckets and fetching the object.

Another option may be to use Minio instead (if that suits your use-case). We have done tests with Minio and it works fine with Sneller.

Disabling the ETag check is not a suitable solution, because it could result in problems when sdb is run in parallel with executing queries. Feel free to disable the check yourself if you are certain that you never run it in parallel.

[rklra]

Hi, thanks so much for the info;

I will try to get the changes made to ceph but while waiting what would be the best interim solution to this issue without causing any execution issues?

Unfortunately using aws/minio is not possible for our use case (we have an extremely write heavy workload that makes these solutions infeasible to operate, but is seemingly well suited for sneller).

Thanks again for the help with resolving these issues!

[ramondeklein]

The safest way would be to interleave sdb and Sneller query execution (if that's feasible).

You also may want to check if it's feasible to change the code to check timestamps instead of etags, but it's not a something that is currently supported. Also don't know if that would be completely safe, because timestamps are often not very precise.

[philhofer]

The ETag checking exists to protect the query engine against binary data that wasn't produced by our ingestion engine. In the cloud platform, this was important to help protect us against users tampering with data that we put in their buckets. The ETag checking does not exist to protect us against data races; we use unique (randomly-generated) object names for everything written by the ingestion engine, so we'll never deliberately overwrite a packfile. If you are self-hosting and your threat model does not involve people having write access to your S3 buckets, then it's fine to eliminate the ETag verification. Just keep in mind that the query engine assumes the data being read from S3 is trusted, and without the ETag verification (ultimately rooted in the MAC of the index file) there is no cryptographic assurance that a properly authenticated party wrote the packfiles.

…

On Mon, Jan 8, 2024 at 11:30 AM Ramon de Klein ***@***.***> wrote: The safest way would be to interleave sdb and Sneller query execution (if that's feasible). You also may want to check if it's feasible to change the code to check timestamps instead of etags, but it's not a something that is currently supported. Also don't know if that would be completely safe, because timestamps are often not very precise. — Reply to this email directly, view it on GitHub <#10 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWN7ZU7EOJ7P2B425JNZCDYNRCNTAVCNFSM6AAAAABBQV4PQSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBRGY4TQNJXGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>