license: cc0-1.0
pretty_name: USCIS DACA and TPS application records
USCIS DACA and TPS applications in DuckDB
This snapshot packages the Deportation Data Project's processed USCIS FOIA application records for reproducible analysis in Python, R and SQL. It contains 11,206,545 application records, not unique people or a count of current beneficiaries. The pinned source has observations through February 2026; the November 3 UTC build date is not a newer observation period.
| Table | Application records | Original source fields |
|---|---|---|
daca_applications |
4,526,373 | 31 |
tps_applications |
6,680,172 | 30 |
The full database is 391,131,136 bytes. SHA256: e0957fb5726b990f694b08c86c0bf8f1d785a5b41d5ad5079ecaac2dfe5a2f9f. It retains every original row, field, NULL and duplicate flag, adds a table-local _source_row, and includes _metadata, _columns and _provenance. The full build and released Python/R client checks have passed. The separate catalog integration must be merged before uscis works as a public Datapond identifier. Direct download and local queries are available independently of that integration.
Provenance and license
The publisher repository is pinned to commit dfe86ad18fd04ed6ed8281366d8bfb9c56716c1b. Both complete Parquet files matched their immutable URLs, full SHA256 hashes and combined 208,298,935 bytes before ingestion. source-lock.json records all pins, source schemas and exact NULL baselines; uscis.manifest.json records the full build and measurements. The processed record counts come from actual pinned files, rather than rounded release-announcement totals. The publisher removes fully redacted rows and entirely blank/redacted columns, repairs ELIS UTF-8, harmonizes gender labels and TPS form-type labels, and adds duplicate flags and DACA reconstruction. This importer preserves that processed output; it does not reconstruct excluded agency records.
Sources: publisher catalog, codebook, and pinned source repository. The processed data are distributed under CC0-1.0; the publisher's text is retained in LICENSE-DATA. Cite USCIS data processed by the Deportation Data Project, this source commit and the snapshot date. Updates depend on FOIA releases; no fixed monthly cadence is promised.
Interpretation
Applicants can have multiple applications and repeated administrative records. Approved application rows do not establish how many people currently hold DACA or TPS. There is no verified cross-form, cross-release, ICE or EOIR identity link. _source_row is a one-based ordinal within this exact processed table, not a stable person or application identifier.
| Full-source measurement | DACA | TPS |
|---|---|---|
duplicate_drop_row = true, retained |
149 | 2,603 |
| Missing applicant identifier | 19,135 | 604 |
Unknown duplicate_likely flag |
19,135 | 612 |
| Missing receipt date | 0 | 8 |
| Decision date before receipt date, retained | 15,853 | 297,594 |
Unknown duplicate flags are not false flags. Publisher duplicate recommendations are preserved, with no automatic deduplication. DACA form_type_filled_in copies form_type for ELIS and missing-ID rows; only C3 rows with an ID are reconstructed. Their cumulative approval count, ordered by decision date and including the current row, labels counts at most one Initial and larger counts Renewal (unknown counts stay NULL). Duplicate approvals count; denied/pending renewals after a first approval can be labelled Initial. See the field guide for the exact rule and limits. Source systems differ in status labels, coverage and missingness. Country of birth and processing offices do not provide residence geography. TPS maps failed C3 country-code lookups to Unlisted Entry (Any foreign country not listed), so stored NULL counts do not capture all underlying missingness; DACA uses NULL for failed C3 lookups. Unusual dates, including early LPR dates, remain uncorrected; assess them before calculating durations or cohort outcomes.
The historical manifest key null_duplicate_flags counts only NULL duplicate_drop_row; its value zero does not contradict the 19,135/612 unknown duplicate_likely values in null_counts. A false drop flag on these unknown rows comes from coalescing, not verified absence of duplicates.
FIELD-GUIDE.md explains these limits. DICTIONARY.md defines all 63 table-specific stored fields and shows exact missing counts, including the added ordinal in each table. Build/download timestamps, snapshot status, receipt dates and decision dates have different meanings.
Python and R queries
After downloading the audited uscis.duckdb snapshot, query it locally:
import duckdb
con = duckdb.connect("uscis.duckdb", read_only=True,
config={"threads": "1", "memory_limit": "128 MiB"})
try:
print(con.sql("SELECT data_source, count(*) AS application_rows FROM daca_applications GROUP BY data_source ORDER BY data_source").fetchall())
finally:
con.close()
library(DBI)
con <- dbConnect(duckdb::duckdb(), dbdir="uscis.duckdb", read_only=TRUE)
tryCatch({
dbExecute(con, "SET threads=1")
dbExecute(con, "SET memory_limit='128 MiB'")
print(dbGetQuery(con, "SELECT data_source, count(*) AS application_rows FROM daca_applications GROUP BY data_source ORDER BY data_source"))
}, finally=dbDisconnect(con, shutdown=TRUE))
This example returns 1,326,378 C3 rows and 3,199,995 ELIS rows in the DACA table. Both released Datapond clients passed the same query and full counts via a temporary loopback registry: Python Datapond 0.1.5/DuckDB 1.5.6 and R Datapond 0.1.0/DuckDB 1.5.5. Their actual datapond.connect("uscis") and dp_connect("uscis") calls are in the smoke scripts; a public catalog entry is required before those identifier-based calls work without the test registry.
Reproduce and refresh
Clone the source repository for build scripts and offline fixtures. Use Python 3.11+ and the pinned duckdb==1.5.6 dependency. No Arrow, spatial extension or credentials are required for ingestion.
python -m pip install -r requirements.txt
python acquire_sources.py
python acquire_sources.py --download
python build_database.py --lock source-lock.json --output output/uscis.duckdb
The build uses one thread, 2 GiB DuckDB memory, a 4 GiB spill cap within its private workspace, an 8 GiB work cap and a 2 GiB free-space reserve. Initial acquisition/build preflight requires about 10.2 GiB free. Existing verified sources are reused. Outputs, WALs and manifests are never overwritten; use a new output path. The filesystem must support atomic hard links. Failed work is retained for inspection. The successful work directory retains a hard link to the output. Rebuilds have new _metadata.built_at timestamps and need not have the same file SHA256; compare source pins, schemas and all typed row values. The published checksum identifies the distributed artifact. The measured full build took about 99 seconds with sampled peak process RSS about 1.34 GiB; these observations are not guaranteed peak bounds on other machines.
The importer verifies complete source hashes, schemas, row counts, every typed source value and ordinal in bounded batches, exact NULL baselines and the publisher's duplicate-flag relationship. Missing IDs, source-key duplication and date anomalies are reported without silent repair. VALIDATION.md distinguishes full execution, independent review and client checks. CLOUD-HANDOFF.md documents the bounded workflow and included pilot fixtures.
To refresh, select a new immutable publisher commit, update the complete source pins and expected schema/NULL counts, inspect processing-code changes, and repeat the build, field checks, Python/R queries and independent audits. No automatic refresh or current legal-status inference is supplied.
Snapshot files and Datapond clients
Download the database, browse its versioned files, or inspect the source repository. Verify the full SHA256 above. The catalog PR pins an immutable HF revision; main can later refer to a newer snapshot.
After the separate catalog entry is merged, the released clients use:
import datapond
con = datapond.connect("uscis")
try:
print(con.sql("SELECT data_source, count(*) AS application_rows FROM daca_applications GROUP BY data_source ORDER BY data_source").fetchall())
finally:
con.close()
library(datapond)
con <- dp_connect("uscis")
tryCatch(print(DBI::dbGetQuery(con, "SELECT data_source, count(*) AS application_rows FROM daca_applications GROUP BY data_source ORDER BY data_source")), finally=dp_disconnect(con))
Before catalog integration, use the downloaded snapshot with the local Python/R examples above. The smoke scripts test the same released clients, aggregate query and exact full database through an isolated registry.