seagrep

seagrep searches S3 buckets with regular expressions. It works like grep, but instead of scanning your objects on every query, it builds a trigram index and a compressed snapshot of the decoded content once, stores both in S3, and answers queries from those alone. A typical search over 25,000 objects returns in about 100 ms. The CLI follows ripgrep: same flags, same exit codes, same --json output.

Dual-licensed under MIT or Apache-2.0.

seagrep index s3://my-logs/prod                # build the index, once
seagrep 'req-7f3e9a2c1b' s3://my-logs/prod     # then grep it
seagrep -i 'timeout' s3://my-logs -g '*.gz' -C2 --since 6h

Installation • Usage • Performance • Architecture • Changelog

Why use seagrep?

S3 has no grep. The usual workarounds scan: downloading everything and running rg pays for every object on every query, and Athena bills per byte scanned. seagrep pays the scan once, at index time. After that:

• Queries read small index ranges plus only the snapshot bytes of candidate documents. A pattern that can't match anything answers in microseconds, without a single network request.
• Results are exact. The index only narrows the candidate set; a real Rust regex over snapshot bytes decides every match, so there are no index approximations and no false positives. Matching semantics are verified differentially against ripgrep (scripts/rg-parity/), every index range is SHA-256-verified at query time, and a search over a bucket containing undecodable objects says so instead of silently skipping them.
• Compressed objects (gzip, zstd, bzip2, xz, lz4, snappy, brotli, zlib) decompress transparently, including the multi-member concatenations that ALB and CloudTrail actually deliver.
• Columnar files are greppable. Parquet, Avro, ORC and Arrow rows are projected to canonical JSON lines, and ZIP/TAR members are searched as individual documents.
• Indexing is incremental. Re-runs fetch only new or changed objects, deletions disappear from results immediately, and an unchanged bucket costs one listing.
• --since 6h scopes a search by the timestamps embedded in object keys. It understands 2026/06/09 paths, hive partitions, dt=/date= prefixes, and ALB/CloudTrail/CloudFront filename stamps.
• It speaks to anything S3-compatible: AWS, MinIO, Cloudflare R2.
• Search keeps working after the source objects are deleted, because it only reads the snapshot.

Installation

Prebuilt binaries for Linux (x86_64, arm64), macOS (Intel, Apple Silicon), and Windows ship with every GitHub release:

cargo binstall seagrep   # fetches the prebuilt binary for your platform
cargo install seagrep    # or build from source (Rust 1.94.1+)

Release archives include SHA-256 checksums and GitHub build-provenance attestations. Verify one with gh attestation verify <archive> -R TalkingComputers/seagrep.

Usage

The shape is seagrep PATTERN TARGET, where TARGET is s3://bucket[/prefix]. Credentials come from the standard AWS SDK provider chain, so environment variables, shared profiles, IAM Identity Center (SSO) sessions, credential_process, and container or instance roles all work as usual, and temporary credentials refresh automatically.

First build the index. It lives in the bucket, under <prefix>/.seagrep/ by default, and searches find it automatically: at the searched prefix, at any parent prefix (an index built at s3://b/logs serves a search of s3://b/logs/2026/07, scoped to that subtree), or at a location remembered from an earlier --index run on the same machine:

AWS_PROFILE=my-sso seagrep index s3://my-log-bucket/prod --region us-east-2

Then search:

seagrep 'level":"ERROR' s3://my-log-bucket/prod --region us-east-2

Most rg flags do what you expect. -i/-S for case, -C for context, -w for word boundaries, -F for fixed strings, -l, -c, -m, -q, and --json emits rg-compatible JSON Lines:

seagrep -i 'timeout' s3://my-logs -C2 -g '*.gz' -g '!debug/*'
seagrep -w -F 'foo(' s3://my-code-bucket -l
seagrep 'req-[0-9a-f]+' s3://my-log-bucket --json | jq .

Exit codes are rg's: 0 match, 1 no match, 2 error. Patterns are line-oriented like rg: ^ and $ anchor at every line, and a literal \n in a pattern is an error. To search for a pattern that collides with a subcommand name, use -e: seagrep -e index s3://bucket.

--files lists every indexed key without a pattern (honoring the same scoping flags), so you can see a corpus's shape before searching it:

seagrep --files s3://my-logs -g '*.gz' | head

Searches can be scoped by key or by time:

seagrep 'ERROR' s3://my-logs --since 6h
seagrep 'ERROR' s3://my-logs --since 2026-06-09 --until 2026-06-10 --key-prefix prod/

--key-prefix prunes whole index segments before any fetch, and --key-regex filters keys by pattern. --since/--until take absolute dates or relative 30s/15m/6h/2d/1w values. Keys without a recognizable timestamp are searched anyway, with a note on stderr, so time scoping never silently hides data.

To keep the index fresh, watch mode repeats the listing/diff/swap cycle on an interval, finishes the active cycle cleanly on SIGINT/SIGTERM, and with --json emits tagged indexed, error, and stopped lines on stdout:

seagrep index s3://my-log-bucket/prod --watch --interval 30

If the source bucket is read-only (or you just want the index elsewhere), put it in its own bucket. Pass the same --index location when searching; --index-region and --index-endpoint configure that connection separately:

seagrep index s3://my-log-bucket/prod --index s3://my-search-index/prod
seagrep 'ERROR' s3://my-log-bucket/prod --index s3://my-search-index/prod

For MinIO, R2, or any other S3-compatible store, point --endpoint at it. --concurrency (default 750) caps parallel requests.

Flag summary:

-e PATTERN        multiple patterns, OR              -n / -N      line numbers on/off
-F                fixed strings                      --column     1-based match column
-i / -S / -s      ignore / smart / sensitive case    --heading    group under key (tty default)
-w                word boundaries (rg half-bounds)   --no-heading key:line:text (pipe default)
-l                files with matches                 -g GLOB      include/!exclude key globs
-c                count matching lines               -q           quiet, exit at first match
--count-matches   count individual matches           --color WHEN auto/always/never/ansi
-m NUM            max matching lines per object      --json       rg-compatible JSON Lines
-A/-B/-C NUM      context lines with -/-- separators --stats      candidate stats to stderr

Object formats

Format detection is magic-first: extensions are not trusted, with one exception. Brotli and zlib have no reliable container magic, so only .br, .zlib, and .zz select those decoders, and the entire stream must validate.

Projection and decompression happen at one canonical decoder boundary, so the index and the verifier see identical bytes. Truncated or corrupt-tailed streams salvage: the cleanly decoded prefix is searched and a warning names the object. Undecodable objects are excluded loudly, never silently searched incorrectly.

A few deliberate rejections: raw (unframed) snappy has no magic bytes and is undetectable by design, so it's unsupported as an object format (it still decodes fine inside Avro files, where the container names the codec). lz4 legacy frames (lz4 -l output) are detected and rejected loudly rather than decoded. Expansion is capped at 64 GiB per physical source, 100,000 archive members, and four nested format layers; oversized decoded output spills to private temporary files instead of memory.

How the index and query pipelines work — crate boundaries, segment format, memory bounds — is covered in ARCHITECTURE.md.

Performance

Real corpora on real S3 (us-east-2; index built once, timings are full process wall time including credential handling):

One caveat worth knowing: candidates are whole documents, so a common token inside very large objects (multi-GB gzip, 100 MB parquet shards) degrades to decoding those objects — seconds, not sub-second, the same work ripgrep would do. Block-level candidates are the planned fix.

Numbers from the tracked benchmark: 25,000 synthetic 4 KiB objects on MinIO, release build, three measured iterations after one warmup. Every corpus, planted hit count, candidate count, and byte count is deterministic and checked before timing. Reproduce with make bench-minio BENCH_OBJECTS=25000 BENCH_ITERATIONS=3.

CI reruns the end-to-end benchmark and a microbenchmark suite (make bench-micro) for every pull request, gates statistically confident regressions, verifies exact hit counts, and enforces peak-RSS ceilings across large-object, archive, and churn workloads. The committed benches/baseline.json is the reporting reference; refresh it only from CI's bench-micro artifact.

As always with benchmarks: this is one corpus with one object-size distribution on local MinIO. Your latencies against real S3 will include network round-trips; the shape (pruning ratio drives cost) is the durable part, not the exact milliseconds.

Security

Use private buckets. The default index lives under <source-prefix>/.seagrep/, and --index can place it in a separately permissioned bucket. The index contains compressed canonical decoded content, not only grams, so protect it with the same access controls, retention policy, and encryption requirements as the source data. seagrep contacts the configured source and index S3 endpoints plus the AWS credential endpoints required by the active SDK provider chain, and nothing else.

Report vulnerabilities privately; see SECURITY.md.

Contributing

Read ARCHITECTURE.md before changing index, query, or S3 behavior, and CONTRIBUTING.md for setup and the CI checks. The differential test suites are the correctness contract: indexed search must exactly equal a decoded full scan, for every format, both gram strategies, and every index lifecycle state.

License

Licensed under either of:

• MIT license (LICENSE-MIT)
• Apache License, Version 2.0 (LICENSE-APACHE)