Skip to content

Benchmark · Vendor-run telemetry rehearsal · Separate historical storage tests · Larger recovery campaign unfinished

8DB storage | Keep working when the connection disappears

16,384 records captured offline. 16,512 verified after delivery and reopening. Explore what protected local storage can do for an application that must keep working.

Published
Reading time7 minutes
StorageOfflineDurabilityOmnimodal

Your application still has a job when the network goes away. It needs to save the next observation, protect what it already holds, and get the right information to the right recipient when the connection returns.

In a completed local telemetry rehearsal, 8DB captured 16,384 records while disconnected, then delivered them alongside 128 new records. A separate verifier reconciled all 16,512 records, their exact content and signed receipts. A fresh process reopened the stores and confirmed the expected data with no pending deliveries. This was a vendor-run Linux rehearsal using local TCP endpoints and test keys. Rehearsal evidence

That is the storage story worth evaluating: capture, protection, delivery and recovery working together. 8DB is an embeddable, omnimodal database, so that foundation can serve an application that needs more than one kind of data operation.

Follow the work, not just the write

Completed local rehearsal: 16,384 disconnected records plus 128 new records, delivery with signed receipts, then fresh-process recovery; all 16,512 records reconciled.

The useful question for a field application is whether saved work remains usable throughout its journey. For a sensor system, that could mean collecting observations offline and delivering a verifiable backlog. For an inspection application, it could mean keeping a service record connected to its location and history. The telemetry rehearsal tests the first workflow; the richer applications are deployment opportunities to qualify.

Fewer moving parts between capture and delivery

8DB brings native data operations, protected capture and verifiable delivery into the same product. For a field application, the practical advantage is less work connecting the database to the rest of the job.

Customer needWhat 8DB bringsHow the alternatives compareWhy it matters
More than key-value storageNative time-series, graph, spatial and analytical operations alongside storageRocksDB, LMDB and redb provide key-value foundations; richer operations require application logic or additional enginesPotentially fewer stores and copies to connect, secure and recover
Proof that the right records arrivedThe completed telemetry rehearsal checked exact content, signed delivery receipts and a fresh-process reopenSQLite / SQLCipher provide local storage and encryption; the application must supply or integrate this delivery-and-verification workflowLess delivery and audit plumbing for the application team to build
Useful work without a connection16,384 offline records delivered with 128 new records in the completed rehearsalCouchbase Lite also supports offline work and direct peer sync. Here, compare 8DB's native data operations and required protection with the complete Couchbase deploymentKeep collecting locally, then reconcile the backlog; offline sync alone is shared capability
Data that stays useful on the deviceAn embeddable engine with several native query modalitiesSQLite / SQLCipher offer embedded SQL; Couchbase Lite offers embedded document queries and sync; the key-value engines need application-level query logicMore ways to use local context before uploading it or assembling another service

Capability and workflow comparison. Fewer integrations, copies and recovery procedures are architectural benefits to evaluate, not measured cost savings. The completed telemetry evidence covers the stated rehearsal; the larger interruption test below remains unfinished. Sources: SQLite, SQLCipher, RocksDB, LMDB, redb and Couchbase Lite peer synchronization.

For an inspection team, that could mean keeping an observation, its location and its equipment history together on the device, then delivering a verifiable record when connectivity returns. 8DB's advantage to evaluate is how much of that complete application one engine can cover. The native performance articles linked below provide evidence for individual modalities; this richer inspection workflow is a deployment opportunity, not a completed benchmark.

More than a day of actual disconnected capture

A subsequent campaign completed 864,000 locally acknowledged records over 24 hours, 9 minutes and 52 seconds. That was elapsed execution, not an accelerated clock simulation. A WSL kernel panic interrupted the later reconnect stage, so full payload recovery and delivery verification remain unfinished for that larger campaign. Recorded outcome

The selected security profile intentionally refused protected access after the boot identity changed. Work is underway on fresh, signed recovery authorization controlled by the deployment owner, including an offline approval exchange. Its first scope is verification and export; it must not revive expired service permissions. This is an accepted design under implementation, while the completed smaller rehearsal remains a separate recovery result. Recovery decision

Faster reopening and less repeated writing

The lower layers contribute their own opportunities. In a historical Apple M1 test, the Zero-WAL path opened a store with 50,000 committed records in 0.098 ms, versus 110.6 ms for 8DB's own WAL path: about 1,100× faster for this warm, clean storage-manager reopen. It did not time a complete application restart or recovery after physical power loss.

Historical internal comparison: warm clean reopening of 1,000, 10,000 and 50,000 committed records, Zero-WAL versus 8DB WAL.

A newer native section-batching candidate reduced application-issued payload writes by 31.8%, from 33.83 MB to 23.08 MB. All three matched pairs retained 109 commits and 109 WAL syncs, with fresh-reopen checks. The synthetic capture-shaped fixture used strict WAL with encryption off; production adoption gates remain open. This measures fewer bytes issued by the application, not physical SSD wear or energy savings.

Native section batching: 33.83 MB versus 23.08 MB application-issued payload writes, a 31.8 percent reduction at the same 109 commits and 109 WAL syncs.

For battery-powered devices and nonreplaceable storage, avoiding unnecessary work could matter as much as peak speed. The byte reduction is measured. Its effect on battery life, physical writes and device lifetime needs instrumented testing in the intended deployment.

Keep the application responsive while saving

A separate M1 mixed-load campaign recorded 2.44–3.40 ms read p99.9 across three valid runs while offering 1,000 writes per second, with durable commits every 100 writes. Across 1,603,273 verified reads, the retained campaign record reports zero value errors. These are internal mixed-load results, not a matched competitor ranking; they measure batched durability rather than one durable commit per write.

That is a useful direction for a field application: collect new work while the user can still consult the history. The store's batching and persistence choices belong in that decision alongside its fastest standalone benchmark.

Where the single-write results fit

Single-operation latency still matters. In the June M1 campaign, SQLite FULL with macOS fullfsync enabled recorded 5.08 ms per commit, compared with 10.50 ms for 8DB SyncNow. SQLite was faster in that historical configuration. Later 8DB SyncNow observations were 5.54–5.64 ms, but the contemporary RocksDB reference used different keys, values and timer boundaries, so it cannot establish a precise ranking.

Those observations belong beside the recovery, mixed-load and workflow results. A single durable PUT does not price the entire application. The next matched comparison will include equivalent protection and completed work, plus first useful query, delivery, recovery, memory and background maintenance.

One storage foundation, many native capabilities

An inspection record, a time series, a location and a relationship may all belong to the same application. 8DB's native modalities create an opportunity to reduce the number of stores, copies and recovery procedures that application needs. The savings depend on the actual replacement architecture and remain to be measured.

Our separate articles cover time-series ingestion, key-value performance, graph lookups and analytics. Each has its own measured scope; their headline results are not added together into an untested combined claim.

Bring us the work your application must preserve, the query it must keep answering and the interruption it must survive. Plan an evaluation around that complete job.

Evidence notes

This article combines separately identified 8Braid engineering campaigns, not a single universally qualified build. The telemetry source is a96ab656; the warm-reopen source is 366e87e4; the batching candidate is 450940d4 against b271c198. Mixed-load results are retained Airtable campaign evidence; original raw timing files remain a recovery gap. Full source mapping is retained in 8Braid's benchmark archive; the underlying telemetry reports are not public downloads. No physical power-cut, energy, full ACID certification or government approval is asserted.

Continue the technical conversation

Where could this help your work?

Bring a research question, a database workload or an application you want to build. Let’s connect the ideas in this article to an evaluation that matters to your team.

Discuss this work

Continue reading

The next layer of the argument.

Benchmark

Database Performance in Context

8DB analytics | How we compare, and what else you get

Historical analytical query results beside ClickHouse, DuckDB, StarRocks, Doris and PostgreSQL references, plus the broader opportunity of an omnimodal engine.

6 min readHistorical server-path timings · Cross-campaign market context · Two selected analytical queriesRead article
Benchmark

Database Performance in Context

8DB graph lookups | 64× Neo4j’s recorded throughput

218,000 lookups per second for 8DB. A best recorded 3,410 for Neo4j. Both on an M1 laptop. A separate one-hop test also put 8DB ahead, reaching 23,300 queries per second.

12 min readHistorical service comparisons · Simplified queries · Different clients and returned-data contractsRead article
Benchmark

Database Performance in Context

8DB spatial | 14.7 ms vs. 140–13,140 ms for GIS tools

Six million locations, recorded GIS-tool comparisons, competitor timeouts and the broader opportunity of an omnimodal spatial engine.

5 min readHistorical kernel and query-path timings · Different measurement boundaries · Selected SpatialBench workloadsRead article