zk-Analytics is a distributed, end-to-end verifiable, privacy-preserving cloud analytics system. It augments an analytics pipeline with lightweight append-only log commitments and zero-knowledge proofs of correct aggregation and query execution, so an external verifier can check reported results without access to raw logs or the provider's infrastructure. Proofs are generated with the RISC Zero zkVM. This repository is the implementation described in the paper "Zero-Knowledge Cloud Analytics."

The pipeline has three logically separated stages (paper §4–5):

1. Online commitment — data_source/. Each data source emits time-series (timestamp, key, value) log entries and incrementally commits to them with a lightweight SHA-256 hash chain (one constant-size update per log). Committed batches are streamed to the provider through a Kafka dispatcher, and each source may publish hash-chain checkpoints to a public transparency log (Trillian — see deploy/trillian/ and docs/transparency_log.md).
2. Offline distributed aggregation — aggregator/. Shared-nothing aggregators consume committed batches from Kafka into a local RocksDB buffer, group them into fixed-size epochs, and run a RISC Zero guest that verifies the source hash-chain and aggregates the epoch into a query-ready summary (per-key samples, histograms, or a Count-Min sketch). Each epoch is chained to the previous one and persisted; its aggregate state is stored in RocksDB or FoundationDB.
3. Offline query + verification — querier/. An HTTP service answers analytics queries over a time window, returning the answer plus a RISC Zero proof that the window's epochs are authentic (epoch-chain + commitment checks) and that the query was executed correctly. Verifiers validate the proof offline, without raw logs or provider infrastructure.

# RocksDB bindings need clang/libclang:
sudo apt-get update
sudo apt-get install -y clang libclang-dev

# Optional features: Kafka (rdkafka, cmake-build) and Trillian (protoc):
sudo apt-get install -y cmake libssl-dev pkg-config protobuf-compiler

# RISC Zero toolchain (guest compiler + r0vm):
curl -L https://risczero.com/install | bash && rzup install

cargo build --release        # host crates + RISC Zero guest ELFs

Proof generation uses AVX-512 for performance; proof verification does not. FoundationDB 7.1 is required only for the FDB-backed (--features fdb) path.

• CPU: x86-64 with AVX-512; many cores recommended (the zkVM prover scales with core count).
• RAM: ≥ 64 GB — the zkVM prover peaks around 9–10 GB per aggregation/query node.
• Storage: NVMe SSD (RocksDB / FoundationDB backing store).

Each service is a Cargo binary. End-to-end, committed batches flow data_source → Kafka → aggregator → RocksDB/FoundationDB → querier.

# Aggregator: consume a Kafka topic into a local RocksDB buffer and prove epochs
# of type samples | histogram | cm (add --features fdb to store aggregates in FDB).
cargo run -p aggregator --release --features kafka -- --mode samples

# Data source: stream events as a Kafka producer (per-source SHA-256 hash chain).
cargo run -p data_source --bin kafka-producer --release --features kafka -- \
  --events 100000 --batch-size 100

# Querier: HTTP query service (default HTTP_LISTEN=0.0.0.0:8082).
cargo run -p querier --release

For a full local run (Kafka + FoundationDB via Docker, orchestrated in tmux):

./scripts/setup/setup_local_e2e.sh --all   # install deps, Kafka/FDB, RISC Zero toolchain
./scripts/eval/run_local_e2e.sh start      # data_source -> Kafka -> aggregator -> FDB -> querier
./scripts/eval/run_local_e2e.sh status

All services use the same data directory:

• ROCKSDB_PATH (default: /mydata/rocksdb)

• aggregator:
  • agg_mode_state (mode-level chain tip for aggregated epochs)
  • agg_epochs (aggregated epochs + merge proof fields + result_commit)
  • agg_cm_struct (structured CM counts/heap + result_commit)
  • agg_hist_struct (structured histogram totals/table + result_commit)
  • verified_epoch_frames (verified-only storage for samples_epoch)
  • verified_samples_struct (structured per-epoch totals + result_commit for samples_epoch)
  • bad_epoch_frames (quarantine for frames that fail verification)

Reset RocksDB storage:

ROCKSDB_PATH=/mydata/rocksdb ./scripts/setup/reset_rocksdb.sh

It polls epoch_frames, verifies each per-source RISC Zero proof (host-side), and:

• For cm_epoch (epoch_type=cm_epoch):
  • merges CM array by element-wise sum
  • merges topk heap by key (sum counts for matching keys)
  • produces a RISC Zero merge proof (the aggr_cm guest)
  • stores structured CM into agg_cm_struct and the SHA-256 result_commit into agg_epochs
  • stores the aggregate into agg_epochs and deletes the original per-source rows for that (epoch_type, sequence)
• For histogram_epoch (epoch_type=histogram_epoch):
  • merges bucket counts by bucket (sum)
  • produces a RISC Zero merge proof (the aggr_histogram guest)
  • stores structured histogram into agg_hist_struct and the SHA-256 result_commit into agg_epochs
  • stores the aggregate into agg_epochs and deletes the original per-source rows for that (epoch_type, sequence)
• For samples_epoch (epoch_type=samples_epoch):
  • verify-only: moves rows from epoch_frames into verified_epoch_frames
  • extracts (out_commit,total_count,total_sum) from the verified proof output, computes a SHA-256 result_commit, and stores into verified_samples_struct (the stored samples table is (key,key_chain_tip,len,sum,occ); key_chain_tip is a SHA-256 chain over that key’s values, preserving per-key order only)
  • the per-source chain_hash is computed by the zkVM guest as SHA-256(SHA-256(chain_prev || TAG_FINALIZE) || out_commit) where out_commit is a commutative sum of SHA-256 digests over (key,key_chain_tip,len,sum) for occupied slots (so cross-key reordering does not change the commitment)

• ROCKSDB_PATH
• INIT_DB=1 to initialize RocksDB (no-op)
• POLL_MS (default: 500)
• MODES (default: 1,2,3)
• MIN_SOURCES (default: 1) for cm/histogram aggregation trigger
• ALLOW_GAPS (default: 0) if 0, only aggregates sequence = last_seq+1 per mode
• PROOF_COMPRESS (default: 0) controls whether aggregator-generated proofs are compressed
• VERIFY_ONLY_BATCH (default: 100) batch size for samples_epoch verify-only

cd zk-Analytics
ROCKSDB_PATH=/mydata/rocksdb INIT_DB=1 cargo run -p aggregator

Recovery semantics and online resharding internals are documented in docs/INTERNALS.md.

querier serves POST /query:

• Loads epochs in a time window from aggregator tables (agg_epochs / agg_* / verified_samples_struct)
• Builds a RISC Zero proof that checks:
  • each epoch struct matches its epoch_commit (SHA-256 commitment recomputed in-circuit)
  • commitments are bound into a window digest (SHA-256 fold)
  • window merge + query result are correct
  • (optional) if SAMPLES_BIND_RAW=1, also recomputes samples_epoch per-key chains from sample_events (ordered by idx) and rejects epochs whose out_commit / table do not match

• HTTP_LISTEN (default: ${QUERIER_IP}:8082)
• ROCKSDB_PATH
• PROOF_COMPRESS=1 to return compressed RISC Zero (Groth16) proofs (default 0 = recursive receipts)

cd zk-Analytics
ROCKSDB_PATH=/mydata/rocksdb cargo run -p querier

Samples sum:

curl -sS localhost:8082/query \
  -H 'content-type: application/json' \
  -d '{"type":"samples_sum","window":"1h"}'

Histogram bucket:

curl -sS localhost:8082/query \
  -H 'content-type: application/json' \
  -d '{"type":"histogram_bucket","window":"1d","bucket":42}'

CM top-k:

curl -sS localhost:8082/query \
  -H 'content-type: application/json' \
  -d '{"type":"cm_topk","window":"5m","limit":20}'

Samples sum by key prefix/suffix (bitmask match):

# Example: suffix match on low 16 bits (mask = 0xffff)
curl -sS localhost:8082/query \
  -H 'content-type: application/json' \
  -d '{"type":"samples_sum_key","window":"1h","key":123,"mask":65535}'

Benchmarking and the non-ZK native baseline are documented in docs/BENCHMARKS.md.