Scheduler database

The scheduler service holds everything about jobs, slices, Compute Groups and workers. Connect with jnu.connect('scheduler', keystore). Every table below is listed with the policy name you pass as table, who can read it, which columns can be changed, and what each column means.

Column types are the ormdb types from the schema. In JavaScript they arrive as:

Schema type	JavaScript value
id, int53, integer, uinteger, float	number
string, string64k, identifier, opaqueId, URI	string
timestamp, autoNow	Date
unixTime	number, seconds since the epoch
ethAddress	Address from dcp/wallet; use String(value) for the hex form and value.eq(other) to compare
finNum	BigNumber; use .toNumber() or String(value)
json	the parsed value, or null
boolean	boolean

“Owner” in the access column means the identity that signs the query, after proxy-key resolution (see Identity and access). “Via job” means the row is reachable only through a job whose owner is you; “via Compute Group” and “via credentials list” work the same way through computeGroups.owner and credsList.owner.

Jobs

jobs

One row per job. Read by the job owner; the owner can change name, description and link and nothing else. Rows can’t be added or deleted through Jnu.

Column	Type	Meaning
id	autoId	Primary key. The other job tables refer to this number as jobId (or as id for the one-to-one tables).
opaqueId	identifier	Public job id, the value of job.id in the Compute API and in the Portal.
status	string	new, estimation, running, paused, finished, cancelled or corrupted.
owner	ethAddress	Identity that owns the job, that is, the root identity behind the key that deployed it.
address	ethAddress	The job’s escrow account.
paymentAccount	ethAddress	Bank account the escrow is funded from.
startTime	timestamp	When the job was created.
endTime	timestamp	When the job was finalized, or null.
lastUpdate	autoNow	Last change to this row.
lastSliceNumber	int53	Index of the last slice. For a job with a fixed input set this is the slice count.
nextSliceNumber	int53	Next slice to hand out. nextSliceNumber - 1 slices have been distributed.
maxSlicesPerTask	integer	Hard cap on slices per task.
absoluteSlicePayment	finNum	Fixed offer per slice, in DCC.
mvMultSlicePayment	finNum	Market-value multiplier. A job deployed with compute.marketValue(f) stores f here and 0 in absoluteSlicePayment. As of 2026-10-03 the market value is a constant 0.001 DCC per slice; a computed market rate is planned.
duplicationLevel	float	1.0 sends each slice once; 2.0 sends every slice twice; 1.5 sends half of them twice.
duplicationType	string	How duplicates are resolved: race or consensus.
flags	uinteger	Bit flags: 1 localExec, 2 open (slices can still be added), 4 workerConsole, 8 greedyEstimation, 16 discrete, 32 useStrict, 64 nsf (insufficient funds).
needCapabilities	uinteger	Worker capabilities the job needs. Bit flags: 1 v8, 2 spidermonkey, 4 browser, 8 offscreenCanvas, 16 fdlibm, 32 webgpu, 64 KVINv9.
requirements	json	Deprecated. Use needCapabilities.
dataStorageType	string	How to read dataStorageDetails: range, pattern or values.
dataStorageDetails	string64k	Range object, URL pattern or values describing the input set.
resultStorageType	string	How to read resultStorageDetails.
resultStorageDetails	string64k	Where results go.
resultStorageParams	json	POST parameters used with resultStorageDetails when it’s a URL.
name	string	Public name, from job.public.name. Changeable.
description	string	Public description. Changeable.
link	string	Public URL. Changeable.
worktime	string	The runtime environment (worktime) the job runs in, for example map-basic or pyodide.
worktimeVersion	string	Semver range for that environment.

jobs_public

The same table, readable by anyone with any identity. Only opaqueId, name, description and link are returned, and opaqueId is the only column a prototype may filter on.

Warning

As of 2026-10-03 an empty prototype is accepted and returns the public fields of every job on the scheduler. This is expected to change; until it does, don’t put anything in a job’s name, description or link that shouldn’t be world-readable.

jobPerfData

One row per job, keyed by id = jobs.id. Read by the job owner. Written by the scheduler on every result, without locking, so values can be slightly off after a crash. Query it with prototype: { id: job.id }; it has no jobId column.

Column	Type	Meaning
id	primaryId	jobs.id.
lastUpdate	autoNow	Last change.
estimationSlicesRemaining	int53	Estimation slices not yet handed out. The job leaves estimation when this reaches zero (default 3 slices).
measuredSlices	integer	Results that have contributed to the averages, including duplicates.
taskCount	integer	Tasks that included at least one slice of this job.
returnCount	integer	Slices given back by workers without being worked on.
retaskedCount	integer	Slices re-issued by the watchdog.
tooSlowCount, noProgressCount, uncaughtCount	integer	Worker-reported failure counts.
sliceCPUTime, sliceGPUTime	float	Exponential moving average per slice, in milliseconds.
sliceInDataSize, sliceOutDataSize	float	Moving average of bytes to and from the worker per slice.
sliceCPUDensity, sliceGPUDensity	float	Moving average of the fraction of wall-clock time spent on the CPU or GPU. Wall time per slice is sliceCPUTime / sliceCPUDensity.
taskDistribution	float	Moving average of the fraction of a task’s nominal duration this job’s slices took, on requests where the job was eligible. Long slices can exceed 1.
*_variance	float	Exponential moving variance of each average above. The square root is a usable standard deviation.
totalCPUTime, totalGPUTime	int53	Running totals, in milliseconds.
totalInDataSize, totalOutDataSize	int53	Running totals, in bytes.

The moving averages use alpha = min(0.5, unreported ÷ total), so a small job’s averages move quickly and a large job’s settle.

jobFinData

One row per job, keyed by id = jobs.id. Read by the job owner. sliceWorkPaymentTotal is DCC paid out so far; sliceWorkPaymentPending is DCC held in escrow for distributed, unfinished slices. The row is recalculated when the job is paused or finished.

Note

Known issue as of 2026-10-03, expected to change: this table returned no row for finished, paid jobs in testing. Treat it as advisory and fall back to offer × completed slices when it’s empty.

activeSlices

Slices that have been handed out and have no result yet. Read by the job owner. Not authoritative for payment; the worker columns record the most recent assignment and aren’t guaranteed.

Column	Type	Meaning
jobId, sliceNumber	id, integer	Which slice.
status	string	new (created, not sent), scheduled (sent), working, overdue (the watchdog thinks it’s late), returned (given back, nobody working on it), tiebreaker, paused.
beginTime	timestamp	When the last assigned worker started it.
lastProgressTime	timestamp	Activation time or the last progress() report. Workers must report progress at least every 5 minutes.
retaskCounter	integer	How many times the slice has been marked overdue.
lastAssignedWorker	opaqueId	Worker id most recently given the slice.
lastAssignedWorkerIdentity	ethAddress	That worker’s identity.

jobResults

One row per completed slice. Read by the job owner; value can be changed (for example to move a result to new storage). value is a URI: a data: URI holding the result for small results, or a URL into the result store on the scheduler for large ones. As of 2026-10-03 the table has no timestamp column, so completion over time has to be sampled by the caller.

Column	Type	Meaning
jobId	id	jobs.id.
sliceNumber	integer	Slice index, from 1.
value	URI	The result, see above.

count({ table: 'jobResults', prototype: { jobId } }) is the number of finished slices.

jobData, jobArguments, workFunctions

The job’s inputs. All read by the job owner; value is changeable on jobData and jobArguments (for example to point a slice that hasn’t started at a new remote URI). workFunctions is read-only.

Table	Columns
jobData	jobId, sliceNumber, value (URI): one input datum per slice.
jobArguments	jobId, argIndex, value (URI): the shared arguments passed to every slice.
workFunctions	jobId, value (URI: the stringified work function or a URL to it), requirePath, modulePath, dependencies (json).

workRejectHistory

One row each time a worker rejected the job outright, for example by calling work.reject() or because a requirement couldn’t be met. Read by the job owner.

Column	Type	Meaning
jobId	id	jobs.id.
sliceNumber	integer	The slice that triggered the rejection.
worker	opaqueId	Which worker.
time	timestamp	When.
reason	string	Short reason, for example EWORKREJECT.
errorMsg	string	The message or the top of the stack.
retries	integer	Retries the worker allowed before rejecting.

computeGroupsJobs_byJob

The computeGroupJobs membership table seen from the job side: jobId, computeGroupId, lastUpdate. Read by the job owner. It gives the numeric ids of the groups the job is in; resolving them to names needs the group’s opaqueId, which this view doesn’t carry.

Compute Groups

All of these are read by the Compute Group owner, through computeGroups.owner.

computeGroups

Column	Type	Meaning
opaqueId	opaqueId	Public group id.
owner	ethAddress	Group owner; has elevated permissions.
name, description	string	Changeable by the owner.
joinKey, joinHash, joinAddress	string	Join credentials: a login, a seeded hash of the secret, and an address whose signature is an alternative. Changeable by the owner.
commissionRate, commissionAccount	finNum, ethAddress	Fraction of each slice payment taken as commission, and where it goes.
deployFee, addJobFee	finNum	DCC charged per deployment and per job added.
deployAccess	string	owner or join: who may deploy into the group.
maxTotalPayment, maxConcurrentJobs, maxConcurrentWorkers, maxConcurrentSandboxes, maxConcurrentCPUs, maxConcurrentGPUs, maxConcurrentEscrow	finNum, integer	Administrative limits; null means unlimited.
creationTime, lastUpdate	timestamp	When the group was created and last changed.

computeGroups_public

Readable by anyone: opaqueId, name and description, filterable only on opaqueId. As with jobs_public, an empty prototype lists every group as of 2026-10-03.

computeGroupJobs

Membership: computeGroupId, jobId, lastUpdate. Read by the group owner.

computeGroupAccountingData

Remaining budgets per group: remainingCPUTime, remainingGPUTime, remainingJobs, remainingInDataSize, remainingOutDataSize.

computeGroupWorkerOrigins

URL origins that workers in the group may contact, with type one of any, fetchWorkFunctions, fetchArguments, fetchData, sendResults. The owner can add, change and delete rows.

computeGroupJobPeriodStats

Per group, per job, per period. This is the only table that records what actually happened in a group over time, and the basis for any market-rate figure.

Column	Type	Meaning
computeGroupId, jobId	id	Which group and which job.
periodStart, periodLength	int53	Period start in seconds since the epoch, and its length in seconds.
taskEligibleCount	int53	Task requests from eligible workers.
noSliceTaskCount	int53	Of those, requests that got no slice of this job.
taskedSliceCount	int53	Slices handed out.
totalEarnings	finNum	DCC earned by the group’s workers on this job in the period.
totalCPUTime, totalGPUTime, totalInDataSize, totalOutDataSize	int53	Totals for the period (milliseconds, bytes).
slice* and slice*_variance	float	Per-slice moving averages and variances, as in jobPerfData, scoped to this group.

totalEarnings ÷ totalCPUTime is the realised rate per CPU-millisecond for that job in that group and period. The cgstat tool on the scheduler host dumps, summarizes and compacts these rows.

computeGroupWorkerPeriodStats

Reserved. As of 2026-10-03 only computeGroupId, workerOpaqueId, periodStart and periodLength exist; per-worker supply statistics are planned here.

Workers

Read and changed by the worker’s registered owner. The tables prefixed w$ and cc$ in the policy are the restricted views a worker’s own proxy key uses; an owner’s key uses the names without the prefix.

worker

Registration: opaqueId (public worker id), owner, identity (the worker’s local key), label (usually the host name), transferLock (true means it can’t be re-registered to another owner), flags (bit 0 marks a Portal worker). The owner has full access.

workerRunInfo

Written at every worker start and on config changes. One row per worker.

Column	Type	Meaning
workerOpaqueId, owner	opaqueId, ethAddress	Which worker and whose it is.
type, os, arch, version, browser, browserVersion	string, json	What’s running: browser or dcp-worker, platform, versions.
hostname, username, localIpAddress, macAddress, pid	string	Host identification.
cpuType, cpuCores, gpuType, gpuCores, ram	string, float	Hardware. ram is in MB.
cpuTargetUtilization, gpuTargetUtilization, maxSandboxes	float	How much of the machine the worker will use.
earningsAccount, earnings	ethAddress, finNum	Where it’s paid and what it has earned since the last reset.
status	string	running or end.
lastStartTime, lastUpdate, lastError, lastErrorTimestamp	timestamp, json	When it last started, last changed, and its last error.

workerTaskInfo

Written on every task fetch: taskTime (last fetch), currentTask (json), serial (ring-buffer position). The averageBusySandboxes* and averageIdleSandboxes* columns exist but are marked as not to be used in the schema as of 2026-10-03.

workerSnapshot

A read-only view joining worker, workerRunInfo, workerTaskInfo and dcpWorkerAppConfig, plus a virtual running column. One query for everything about the owner’s workers.

dcpWorkerAppConfig, workerConfigValue

Worker configuration pushed from the database: application settings (syslogUrl, enableSyslog, enableEventLog, startupDelay, startupDelayType) and individual dcpConfig.worker values (keyPath, value). type is active or template; a null workerOpaqueId means the row applies to all of the owner’s workers. Owner has full access.

workerCreds, credsList, credential

Credential storage for workers, currently compute-group join credentials (type cg-join on the list, cg on the credential, with identifier holding the joinKey and secret the joinHash). Owner has full access. Workers read them through a credCookie.

workerNotifications, registerWorkerDirective

Notifications queued for a worker (event, signature) and instructions for worker registration keys. Owner has full access.

Not exposed

consensusResults, reliableEvents, the vw_* views and the check test table have no policy and can’t be queried: the error is Invalid operation scheduler.jnu.<table>.query.