Real-world gdal vector pipeline examples rendered with render_reactflow(). Every diagram is live: drag to pan, scroll to zoom, use the minimap, and click any node to open the inspector with that step’s full arguments, propagated stream state, and GDAL docs link.
TIGER states: a linear ETL pipeline
A classic download-filter-reshape-write flow, straight from the US Census TIGER shapefiles over /vsizip/vsicurl/. Note the edge badges tracking what changes along the stream: the SQL step reduces the schema to 6 fields, and reproject stamps the CRS.
tiger <- paste(
"gdal vector pipeline",
'! read --input "/vsizip/vsicurl/https://www2.census.gov/geo/tiger/TIGER2025/STATE/tl_2025_us_state.zip/tl_2025_us_state.shp"',
'--input-layer tl_2025_us_state --open-option ENCODING=UTF-8 --open-option ADJUST_GEOM_TYPE=FIRST_SHAPE',
"! filter --where \"STATEFP NOT IN ('02','15','60','66','69','72','74','78')\"",
'! sql --sql "SELECT GEOID AS geoid, STATEFP AS state_fips, STUSPS AS state_abbr,',
'NAME AS state_name, ALAND AS area_land_m2, AWATER AS area_water_m2 FROM tl_2025_us_state"',
"! make-valid",
"! set-geom-type --multi --skip",
"! sort --method hilbert",
"! write --output tiger_states.parquet --output-format Parquet",
'--layer-creation-option COMPRESSION=ZSTD --layer-creation-option GEOMETRY_ENCODING=GEOARROW'
)
render_reactflow(tiger, theme = "dark", height = 620)Tee branching: valid / invalid geometry triage
tee fans the feature stream out into side pipelines – here splitting parcels by ST_IsValid() into separate outputs while the main stream continues to a Parquet write. Branch flows render dashed.
triage <- paste(
"gdal vector pipeline",
'! read --input parcels.gpkg --layer parcels_statefp13',
'! tee',
'--tee-pipeline [ sql --sql "SELECT * FROM parcels_statefp13 WHERE ST_IsValid(geom)" --dialect SQLITE',
'! write --output valid_parcels.gpkg --overwrite ]',
'--tee-pipeline [ sql --sql "SELECT * FROM parcels_statefp13 WHERE NOT ST_IsValid(geom)" --dialect SQLITE',
'! check-geometry ! write --output invalid_parcels.gpkg --overwrite ]',
'! write --output all_parcels.parquet --output-format Parquet',
'--layer-creation-option COMPRESSION=ZSTD'
)
render_reactflow(triage, direction = "LR", theme = "light", height = 540)Config-heavy runs and merged schema chains
A production parcels ETL: eight --config options group into the runtime configuration node, thirty-two one-field-at-a-time set-field-type steps (a GDAL nuance – the command accepts one field per step) merge into a single stacked set-field-type ×32 node, and because the pipeline writes GDALG, the graph attaches the implicit streamed output sink. Click the merged node to see every field -> type coercion.
fields <- c(
source_fid = "Integer64", geoid = "String", state_fips = "String",
county_fips = "String", tax_year = "Integer", num_buildings = "Integer",
num_units = "Integer", year_built = "Integer", num_floors = "Integer",
building_sqft = "Integer64", num_bedrooms = "Integer",
num_half_baths = "Integer", num_full_baths = "Integer",
improvement_value = "Integer64", land_value = "Integer64",
ag_value = "Integer64", total_value = "Integer64", tax_acres = "Real",
calc_area = "Real", sale_amount = "Real", sale_date = "Date",
firm_date = "Date", static_bfe = "Real", elevation_min = "Real",
elevation_max = "Real", elevation_avg = "Real", num_fireplaces = "Integer",
updated = "Date", centroid_x = "Real", centroid_y = "Real",
surface_point_x = "Real", surface_point_y = "Real"
)
parcels <- paste(
"gdal vector pipeline",
"--progress",
"--config CPL_DEBUG=ON --config CPL_TIMESTAMP=ON --config CPL_LOG_ERRORS=ON",
"--config CPL_LOG=logs/state_fips=44.gdal.log --config GDAL_NUM_THREADS=ALL_CPUS",
"--config OGR_GPKG_NUM_THREADS=4 --config OGR_SQLITE_CACHE=4000000",
"--config OGR_SQLITE_SYNCHRONOUS=OFF",
"read --input parcels/LR_PARCEL_NATIONWIDE_FILE_US_2026_Q2.gpkg",
"--input-layer lr_parcel_us --open-option LIST_ALL_TABLES=NO",
'! filter --bbox "-71.907258,41.095834,-71.088571,42.018799" --where "statefp=\'44\'"',
"! sql --sql @sql/gpkg.schema.full.sql --dialect SQLITE",
"! make-valid",
"! set-geom-type --multi --skip",
paste(
sprintf("! set-field-type --field-name %s --field-type %s", names(fields), fields),
collapse = " "
),
"! reproject --output-crs EPSG:4326"
)
render_reactflow(parcels, theme = "dark", height = 720)Pasted scripts just work
The same parser accepts pasted PowerShell (or bash) scripts with line continuations – handy when your pipelines live in .ps1 files:
ps1 <- 'gdal vector pipeline `
read --input "G:/data/states.gpkg" --input-layer states `
! filter --where "STUSPS IN (\'GA\',\'FL\',\'AL\')" `
! reproject --output-crs EPSG:5070 `
! write --output southeast.fgb --output-format FlatGeobuf --overwrite'
render_reactflow(ps1, direction = "LR", theme = "light", height = 320)From diagram back to data
Every diagram above is backed by the same parsed model, so nothing here is render-only. The graph is plain data:
g <- pipeline_graph(triage)
g$nodes[, c("command", "category", "branch_role", "description")]
#> # A tibble: 8 × 4
#> command category branch_role description
#> <chr> <chr> <chr> <chr>
#> 1 read source main Read layer 'parcels_statefp13'
#> 2 tee branch main Split stream into 2 side outputs
#> 3 sql attribute tee Transform attributes with SQL
#> 4 write sink tee Write to valid_parcels.gpkg
#> 5 sql attribute tee Transform attributes with SQL
#> 6 check-geometry inspect tee Flag invalid or non-simple geometries
#> 7 write sink tee Write to invalid_parcels.gpkg
#> 8 write sink main Write to all_parcels.parquet as Parquet
g$edges
#> # A tibble: 7 × 4
#> from to kind badge
#> <chr> <chr> <chr> <chr>
#> 1 n1 n2 main <NA>
#> 2 n2 n3 main <NA>
#> 3 n3 n4 main <NA>
#> 4 n2 n5 main <NA>
#> 5 n5 n6 main <NA>
#> 6 n6 n7 main <NA>
#> 7 n2 n8 main <NA>and the pipeline serializes back to a canonical command line, a formatted shell script, or a GDALG specification via render_command_line(), render_script(), and as_gdalg().