Storage · K3 Tables

Serverless SQL on your bucket.

Serverless HTAP, built in Rust, on every K3 bucket. Run analytical SQL with parallel scans, ACID transactions, and transactional-fresh reads — with no warehouse to run.

Start free Read docs

my-bucket / contracts.sqlk3 tables

run

-- Analytics · distributed scan, parallel partitions.
SELECT region,
       COUNT(*)      AS cnt,
       SUM(amount)   AS total,
       AVG(amount)   AS avg_value
FROM contracts
WHERE signed_at > '2026-01-01'
GROUP BY region
ORDER BY total DESC;

✓3 rows · 4 partitions scanned in parallel · 47 ms

region
cnt
total
avg_value
UK
142
£ 8,419,200.00
£ 59,290.14
EU
 98
£ 5,724,600.00
£ 58,414.29
US
 24
£ 1,442,400.00
£ 60,100.00

Distributed query

Plan once.
Scan partitions in parallel.

The planner opens the Delta target, enumerates partitions, and fans out one Ignite worker per partition. Partial results stream back to a reducer that does sum / count / top-k / welford merges in pure compute.

planner

warehouse_planner

opens Delta · rewrites SQL · fans out

warehouse_query · region=UK3,201

warehouse_query · region=EU4,912

warehouse_query · region=US2,734

warehouse_query · region=APAC2,000

reducer

aggregate_results

sum · top-k · welford

planscan parallel 4 → merge47 ms end-to-end · 12,847 rows aggregated

HTAP

Read-after-write,
without the warehouse.

Writes append to a WAL on S3 — every PUT is a single object, so concurrent writers don't fence each other. A separate compactor drains WAL → Delta in batches. Read with freshness: 'strong' to merge WAL + Delta in the same query — newly-written rows are visible immediately, with no staging or replica lag.

freshness: eventual· Delta only

8 ms

client→warehouse_query→Delta scan

freshness: strong· WAL ∪ Delta

14 ms

client→htap_query→WAL overlay→Delta scan

Serverless functions — planners, workers, reducers, compactors

warehouse_querywarehouse_insertwarehouse_mergewarehouse_plannerwarehouse_aggregate_resultshtap_upserthtap_compact_tickhtap_query

Time travel — every write is a version

Read the table as it was.

Every INSERT, MERGE and OPTIMIZE creates a new transactional snapshot. Query any version with AS OF VERSION or AS OF TIMESTAMP — or roll the whole table back with one statement.

Versions7 of 1,297

v1297headMERGE

89 rows · mohrizk · 2m ago

v1296INSERT

412 rows · mohrizk · 14m ago

v1295OPTIMIZE

system · 1h ago

v1294INSERT

320 rows · mohrizk · 3h ago

v1293SCHEMA

mohrizk · yesterday

v1292MERGE

67 rows · scriptum · 2d ago

v1291VACUUM

system · 3d ago

Reading @v1294· 2026-05-21 11:42:08 UTCsnapshot frozen

SELECT * FROM contracts AS OF VERSION 1294 WHERE region = 'UK'

id
region
signed_at
amount
9912001
UK
2026-04-01
58,200.00
9912003
UK
2026-04-04
91,400.00
9912005
UK
2026-04-07
67,300.00

3 rows·18 ms·0 files rewrittenΔ HEAD shows 4 rows · row 9912007 added in v1296

Roll back the whole table

RESTORE TABLE contracts TO VERSION 1294;

Atomic · ms-level · still reversible

Connect — the table speaks five wires

Plug it into any app.

Use the Python or TypeScript SDK, the dodil CLI, any Postgres-wire client (Metabase, DBeaver, Grafana, Tableau), or the raw TableService gRPC. Same table, same SQL.

connect.python

from dodil_k3 import Client
 
k3 = Client(region="uk", token=DODIL_TOKEN)
rows = k3.tables("my-bucket").query("""
SELECT region, SUM(amount) AS total
FROM contracts WHERE signed_at > '2026-04-01'
GROUP BY region
""")
 
for r in rows:
print(r.region, r.total)

Installpip install dodil-k3 · 3.10+ · async alsosame auth · same SQL · same table

Patterns — built for four shapes

Four patterns. One engine.

Real-time dashboards, audit trails, ML feature stores, streaming sinks — same table type, same SQL. No separate “OLTP” and “OLAP” stacks.

Real-time dashboards

HTAP freshness wins here.

SELECT region, SUM(amount)
  FROM orders
  WHERE created_at > 
    NOW() - INTERVAL '1 hour'
  GROUP BY region

live · /sec1,289 ↑

Strong read · < 14 ms · live

Audit trail

Append-only + time travel.

INSERT INTO audit (actor, action,
  target, at) VALUES (...);
 
-- replay 3 days ago
SELECT * FROM audit
  AS OF TIMESTAMP
    '2026-05-18T00:00Z'

v1290↑ replay v1294v1297

Immutable · auditable · cheap

ML feature store

Low-latency entity lookup.

SELECT * FROM user_feats
  WHERE user_id = $1
 
-- batch warm
MERGE INTO user_feats
  USING training_run_47
  ON user_id;

lookupuser_id = 47291
· 47289 · feats[256] 
· 47290 · feats[256] 
→ 47291 · feats[256]  · 1.1 ms
· 47292 · feats[256] 

1 ms PK lookup · HTAP

CDC sink / streaming ingest

Kafka / Scriptum → table.

MERGE INTO orders AS t
  USING staging AS s
  ON t.id = s.id
  WHEN MATCHED THEN UPDATE
  WHEN NOT MATCHED THEN
    INSERT

events / sec14,200
kafka━━▸MERGE━━▸orders

Exactly-once · WAL durable

Data browser — Excel-compatible

View any table in the UI. Use Excel formulas.

Every K3 table opens as a full spreadsheet in the browser — sort, filter, conditional formatting, pivots, comments, and Excel-compatible formulas like =SUMIFS, =VLOOKUP, =XLOOKUP running live against the table. No CSV export. No separate “data tool”.

contracts · Q2 2026

live · synced 2s ago

142 rows · 7 visible

D8fx=SUMIFS(amount, region, “UK”)

A · id
B · regionUK
C · signed_at
D · amount
E · channel
1
9912001
UK
2026-04-01
58,200.00
web
2
9912002
EU
2026-04-02
42,100.00
partner
3
9912003
UK
2026-04-04
91,400.00
web
4
9912004
US
2026-04-05
24,800.00
web
5
9912005
UK
2026-04-07
67,300.00
api
6
9912006
EU
2026-04-10
33,750.00
partner
7
9912007
UK
2026-04-12
78,900.00
web
8
UK total
295,800.00

selectionSum295,800.00Count4Avg73,950.00Min58,200.00Max91,400.00region = “UK” · 4 of 7 rows match

FormulasFilters + sortConditional formattingPivot tablesCell commentsNamed ranges

Pricing

Pay per query.
Not per cluster.

Storage at the K3 Objects rate. Compute (planner + workers + reducer) at the Ignite vCPU + GB-RAM rate, charged per second of execution. No idle warehouse.

$0.0199

/ GB · month

Delta files in S3

$0.0432

/ vCPU-hour

Query compute

$0.0060

/ GB-RAM-hour

Query compute

Free

PUT · GET · LIST

no per-request charge

See pricing

Why K3?

Knowledge in three layers.

K3 = Knowledge 3. Tables is the structured-rows layer — extract facts from your objects and query them with SQL, all in the same bucket as the raw files and the embeddings.

Objects

Raw bytes. The S3 layer.

Tablesyou are here

Rows extracted from those bytes. Delta + DuckDB.

Vectors

Embeddings of the same content. Hybrid search.

FAQ