Zarr file over HTTP fetches several times the same time axis variable

[raphaeljolivet]

What happened?

When loading a Dataset from Zarr over HTTP the time chunk is fetched 3 times from the remote location.

Example Zarr dataset (on google bucket) :
https://storage.googleapis.com/test_zarr_oie/IEA_PVPS.zarr/ADR/zarr.json

What did you expect to happen?

I expect the following URL to be fetched once only :
https://storage.googleapis.com/test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0

Instead it is fetched 3 times (logs from mitmproxy) :

16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/.zattrs                                                       404 application/xml  201b 358ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/.zgroup                                                       404 application/xml  201b 384ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/.zmetadata                                                    404 application/xml  204b 346ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/zarr.json                                                     200 …plication/json 40.5k 204ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0                                                      200 …n/octet-stream ….04k  26ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0                                                      200 …n/octet-stream ….04k  24ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/station_name/c                                                200 …n/octet-stream   27b  31ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0                                                200 …n/octet-stream ….04k  25ms 

Minimal Complete Verifiable Example

# To spy the request, you may setup a local proxy first : 
# `mitmproxy --mode reverse:https://storage.googleapis.com/ --listen-port 1111`

# Initial URL
# URL="https://storage.googleapis.com/test_zarr_oie/IEA_PVPS.zarr/ADR"

# Proxy URL 
URL="http://localhost:1111/test_zarr_oie/IEA_PVPS.zarr/ADR"

ds = xr.open_dataset(URL, engine="zarr")
print(ds)

MVCE confirmation

• Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• Complete example — the example is self-contained, including all data and the text of any traceback.
• Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• New issue — a search of GitHub Issues suggests this is not a duplicate.
• Recent environment — the issue occurs with the latest version of xarray and its dependencies.

Relevant log output

16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/.zattrs                                                       404 application/xml  201b 358ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/.zgroup                                                       404 application/xml  201b 384ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/.zmetadata                                                    404 application/xml  204b 346ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/zarr.json                                                     200 …plication/json 40.5k 204ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0                                                      200 …n/octet-stream ….04k  26ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0                                                      200 …n/octet-stream ….04k  24ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/station_name/c                                                200 …n/octet-stream   27b  31ms 
16:28:07 HTTPS GET    storage.googleapis.com /test_zarr_oie/IEA_PVPS.zarr/ADR/time/c/0

Anything else we need to know?

No response

Environment

[TomNicholas]

Thanks for reporting this. I'm not familiar with mitmproxy.

A useful next step would be to try to reproduce these logs using the zarr.storage.LoggingStore, something like this:

from obstore.store import HTTPStore
from zarr.storage import ObjectStore, LoggingStore

httpstore = HTTPStore.from_url(URL)
remotestore = ObjectStore(httpstore)
loggingstore = LoggingStore(remotestore)

ds = xr.open_zarr(URL)
print(ds)

[raphaeljolivet]

Here you go :

import xarray as xr
from obstore.store import HTTPStore
from zarr.storage import ObjectStore, LoggingStore
import logging
import warnings
warnings.filterwarnings('ignore')


URL="https://storage.googleapis.com/test_zarr_oie/IEA_PVPS.zarr/ADR"


class GrepLogHandler(logging.StreamHandler):
    def __init__(self, pattern, stream=None):
        super().__init__(stream)
        self.pattern = pattern

    def emit(self, record):
        message = self.format(record)
        if self.pattern in message:
            super().emit(record)

logger = logging.getLogger()
logger.addHandler(GrepLogHandler("Calling ObjectStore.get"))

httpstore = HTTPStore.from_url(URL)
remotestore = ObjectStore(httpstore)
loggingstore = LoggingStore(remotestore)

ds = xr.open_zarr(loggingstore)

This gives the following output :

Calling ObjectStore.get(zarr.json)
 Calling ObjectStore.get(.zgroup)
 Calling ObjectStore.get(.zattrs)
 Calling ObjectStore.get(.zmetadata)
 Calling ObjectStore.get(time/c/0)
 Calling ObjectStore.get(time/c/0)
 Calling ObjectStore.get(station_name/c)
 Calling ObjectStore.get(time/c/0)

[d-v-b]

why are we fetching .zgroup and .zattrs? is it possible to tell xarray in advance that it's v3 data?

[TomNicholas]

why are we fetching .zgroup and .zattrs? is it possible to tell xarray in advance that it's v3 data?

I think this is because xarray doesn't know in advance which format zarr you are pointing it at. I think that if you pass zarr_format=2/3 to xr.open_zarr it will only look for one or the other metadata keys.

[TomNicholas]

But I have no explanation for why it fetches the time chunk 3 times.

[dcherian]

This happens specifically for time variables when decoding datetimes. We grab the first and last element to check that we can decode to the right dtype without errors. So the first two requests are for these two elements; and then the last one is for the whole array.

In Xarray, we should be able to consolidate this to two requests in total by grabbing the first and last element at the same time. You can avoid this totally by enabling caching at the Zarr layer, using either fsspec or Zarr.