Arqly
Sign in
Build a wide-column store (Cassandra / DynamoDB family) (13 scenes)
Scene 12 · Design canvas — pick every knob
Given a real workload, pick partition key, RF, W/R, vnodes, and DC topology — and grade the result.
Previously

Every primitive is now on the table; the question is which positions of the knobs survive a real workload.

Scene 13
Design canvas: pick every knob
Diagram
A workload tile on the left selects the brief. The center column is six knobs the prior scenes taught — partition-key strategy, RF, W (CL_write), R (CL_read), vnodes-per-node, multi-DC topology. The right column is a per-knob verifier that grades each choice against the chosen workload, with a one-line reason.
WORKLOADTIMESERIES1B events/day, 30-day reten…▸ activeMULTI-REGION KVlow-latency reads in two re…WRITE-HEAVY INGESTsustained write spikes, RF …KNOBSPARTITION KEYnatural·RF3·W (CL WRITE)ALL·R (CL READ)ONE·VNODES / NODE1·MULTI-DCsingle-dc·VERIFIER✗partition keyraw key → unbounded growth; >100 MB partitions, GC paus…✓RFRF=3 tolerates 1 failure at QUORUM; the default everywh…✗W (CL write)ALL stalls on any single replica blip; timeseries needs…✓R (CL read)Timeseries reads tolerate some staleness; ONE is fast.✗vnodes / node1 token / node: lopsided ring; bootstrap streams from O…✓multi-DCSingle-DC fits the workload; no cross-region complexity.Timeseries: 1B events/day, 30-day retention. Punishes wide partitions; rewards bucketed keys + TWCS.
natural key → wide partitions on hot users
CL_write=ALL kills availability →
1 vnode/node → lopsided ring
The defaults — raw partition key, CL_write=ALL, 1 vnode/node — are wrong on purpose for the timeseries brief. The verifier panel is glowing red on the knobs that will hurt first. Read it, then continue to start tuning.
Implementation
Verifier.verify
grade each knob against the chosen workload
1def verify(workload, choices):
2    return [
3        verify_partition_key(workload, choices),
4        verify_rf(workload, choices),
5        verify_cl_write(workload, choices),
6        verify_cl_read(workload, choices),
7        verify_vnodes(workload, choices),
8        verify_multi_dc(workload, choices),
9    ]
Verifier.verify_cl_write
the same W is right or wrong depending on the brief
1def verify_cl_write(workload, c):
2    if workload == 'timeseries':
3        if c.cl_write == 'ALL': return fail
4        if c.cl_write in {'QUORUM', 'LOCAL_QUORUM'}:
5            return ok
6        return warn  # ONE: no read-your-writes
7    if workload == 'multi-region-kv':
8        if c.cl_write == 'LOCAL_QUORUM': return ok
9        if c.cl_write == 'ALL': return fail
10        return warn
11    # write-heavy-ingest
12    return ok if c.cl_write == 'ONE' else warn
Verifier.verify_partition_key
wide-partition guard — bucketed for timeseries, raw for KV
1def verify_partition_key(workload, c):
2    if workload == 'timeseries':
3        if c.partition_key == 'time-bucketed':
4            return ok  # bounded; TWCS-friendly
5        if c.partition_key == 'natural':
6            return fail  # >100MB partitions
7        return warn
8    if workload == 'multi-region-kv':
9        if c.partition_key == 'natural': return ok
10        return warn  # KV access by id
11    # write-heavy-ingest
12    return ok  # composite or time bucketing
PreviousGossip keeps everyone informed