[NewGraphEnvironment]

Second of three independent PRs (see ).

chunk_status records a band image that fails to open, but two other failures leave the chunk OK:

• A pixel read that fails after the image opened (a corrupted tile, or a refused or throttled HTTP range request on /vsicurl/): warp_simple() ignores ChunkAndWarpImage()'s return, so the NaN output is used as data. For a mask image the scene is then used unmasked, since NaN never matches a mask value, and cloud reaches the composite.
• A mask image that fails to open: the image loop continues without set_status(), so the scene is dropped with status OK.

Reproducible example (no network)

gdalcubes 0.7.5 · R 4.5.2 · macOS 15.6.1 (arm64) · GDAL 3.13.0. Two dates, a band and an SCL-style mask; the second date is all cloud, and its mask file is cut short so it opens but cannot be read.

library(gdalcubes)
d = tempfile(); dir.create(d)
tif = function(prefix, date, value) {
  v = cube_view(srs = "EPSG:32609", extent = list(left = 0, right = 400, bottom = 0, top = 400,
                t0 = date, t1 = date), dx = 10, dy = 10, dt = "P1D")
  invisible(write_tif(gdalcubes:::.raster_cube_dummy(v, 1, value), d, prefix = prefix, COG = TRUE,
                      creation_options = list(COMPRESS = "DEFLATE")))
}
tif("B04_", "2021-07-05", 500); tif("SCL_", "2021-07-05", 4)   # clear scene
tif("B04_", "2021-07-15", 900); tif("SCL_", "2021-07-15", 9)   # cloud scene
fmt = file.path(d, "format.json")
writeLines(c('{"description": "x", "tags": ["test"], "pattern": ".*\\\\.tif",',
  ' "images": {"pattern": ".*_([0-9-]{10})\\\\.tif"},',
  ' "datetime": {"pattern": ".*_([0-9-]{10})\\\\.tif", "format": "%Y-%m-%d"},',
  ' "bands": {"B04": {"pattern": ".*/B04_.*"}, "SCL": {"pattern": ".*/SCL_.*"}}}'), fmt)
col = create_image_collection(list.files(d, "\\.tif$", full.names = TRUE), format = fmt,
                              out_file = file.path(d, "col.db"))

# Damage the cloud scene's mask: drop the last bytes, which in a COG are tile data.
# The file still opens; reading its pixels fails.
f = file.path(d, "SCL_2021-07-15.tif")
b = readBin(f, "raw", file.size(f)); writeBin(b[1:(length(b) - 16)], f)

v = cube_view(srs = "EPSG:32609", extent = list(left = 0, right = 400, bottom = 0, top = 400,
              t0 = "2021-07-05", t1 = "2021-07-24"), dx = 10, dy = 10, dt = "P20D",
              aggregation = "median")
out = tempfile(fileext = ".nc")
raster_cube(col, v, mask = image_mask("SCL", values = c(8, 9))) |>
  select_bands("B04") |> write_ncdf(out)
nc = ncdf4::nc_open(out)
table(B04 = ncdf4::ncvar_get(nc, "B04"), useNA = "ifany")
ncdf4::ncvar_get(nc, "chunk_status")

On 0.7.5 the cloud scene is used unmasked (the median of 500 and 900) and the chunk is OK:

B04
 700
1600
[1] 0

With this change: 500 in all 1600 cells and chunk_status 2 (incomplete).

Change

• warp() returns nullptr when Initialize() or ChunkAndWarpImage() fails, so read_chunk()'s existing !gdal_out branches mark the chunk INCOMPLETE (ERROR when incomplete_ok = FALSE), for band and mask images alike.
• A mask image that fails to open sets INCOMPLETE, as the band path does.

One behaviour change to be aware of: an image for which GDAL cannot compute a source window also makes ChunkAndWarpImage() fail, so it now gives INCOMPLETE instead of silent NaN with status OK.

Adds inst/tinytest/test_chunk_status.R (band read, mask open and mask read failures, plus clean controls; 10 of its 32 assertions fail on 0.7.5). The full suite passes. R-CMD-check on master is currently failing as well, independent of this change.

---

Fork review copy. Everything above the line is the draft for the upstream PR to appelmar/gdalcubes; not posted. Replaces the issue drafted in #6, which is folded in here.

Relates to NewGraphEnvironment/drift#99
Relates to NewGraphEnvironment/sred-2025-2026#16

🤖 Generated with Claude Code

[NewGraphEnvironment]

Posted upstream as appelmar#114 (2026-10-14).