> ## Documentation Index
> Fetch the complete documentation index at: https://docs.octavehq.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Google Cloud Storage

> Generate a deployable pipeline that loads GTM events into Octave from Google Cloud Storage

Exports land in a bucket as Parquet, CSV or JSONL, usually partitioned by date. Eventarc turns a finalized object into a trigger, and the blob prefix doubles as a resumable watermark.

## What you need

* Read access to the export prefix
* An Octave workspace API key, from **Settings → Integrations**
* A Google Cloud project to deploy into

## Where this runs

A Cloud Run function on an Eventarc trigger covers the incremental run. For the backfill, a Cloud Run job is the better fit — same container, but it runs for up to 24 hours instead of being bounded by a request.

|                         |                                                        |
| ----------------------- | ------------------------------------------------------ |
| **Runtime**             | Cloud Run function (2nd gen), Python 3.12              |
| **Incremental trigger** | Eventarc on `google.cloud.storage.object.v1.finalized` |
| **Scheduled sweep**     | Cloud Scheduler, hourly or nightly                     |
| **Secrets**             | Secret Manager                                         |
| **Watermark**           | Firestore document, or the blob prefix itself          |
| **Initial backfill**    | Cloud Run job — up to 24 hours per execution           |

## The prompt

Copy this into Claude Code, Cursor or any coding agent. It carries the whole flow —
reading your warehouse, mapping the columns, posting to Octave, and deploying the result.

Fill in section 1 with your real schema and a few sample rows (read the Parquet schema off one blob with `pyarrow.parquet.read_schema()`, or `head -5` a CSV or JSONL object). The more of
your actual data it can see, the less it has to guess.

```text theme={null}
I want to load GTM history from Google Cloud Storage into Octave, and keep it current after that.
Start by working out the column mapping. Do not write any code until the mapping is
agreed.

## 1. Here is my data

Location: gs://my-bucket/gtm-exports/
Format: Parquet, partitioned by dt=YYYY-MM-DD
Prefixes and what they hold:
  opportunities/  -> deals
  email_log/      -> outbound sends and replies

<PASTE YOUR SCHEMA AND 3-5 SAMPLE ROWS HERE, ONE BLOCK PER TABLE>

## 2. Where this needs to run

Runtime: Cloud Run function (2nd gen), Python 3.12
Trigger: Eventarc on google.cloud.storage.object.v1.finalized
Sweep: Cloud Scheduler as a backstop for missed events
Secrets: Octave API key in Secret Manager
Identity: a dedicated service account, not a downloaded key file
Watermark: Firestore, or the dt= blob prefix itself
Constraint: my initial backfill covers several years. Propose a Cloud Run job for
it, sharing the same container image as the function, rather than stretching the
event-driven path to cover it.

## 3. Here is what Octave needs

Four event types. Every event carries a "type" discriminator.

CRM event - a deal:
{
  "type": "crm",
  "eventTimestamp": "2024-02-03T16:00:00.000Z",
  "eventType": "opportunity_created",
  "opportunityId": "opp_12345",
  "opportunityName": "Acme Corp - Enterprise Plan",
  "amount": "50000.00",
  "currency": "USD",
  "stage": "Qualification",
  "stageCategory": "open",
  "accountName": "Acme Corp",
  "crmAccountId": "acct_67890",
  "crmAccountDomain": "acme.com",
  "contactEmail": "john@acme.com",
  "contactEmails": ["john@acme.com", "priya@acme.com"],
  "ownerEmail": "sarah@company.com",
  "closeDate": "2024-03-31T00:00:00.000Z",
  "crmLastModifiedAt": "2024-02-03T16:00:00.000Z",
  "eventId": "<stable id derived from my row>"
}

Email event:
{
  "type": "email",
  "eventTimestamp": "2024-02-03T10:30:00.000Z",
  "eventType": "sent",
  "subject": "Follow up on our conversation",
  "body": { "text": "Hi John, just wanted to follow up..." },
  "from": { "email": "sarah@company.com", "name": "Sarah Johnson" },
  "to": [{ "email": "john@acme.com", "name": "John Smith" }],
  "crmOpportunityId": "opp_12345",
  "eventId": "<stable id derived from my row>"
}

Call event:
{
  "type": "call",
  "eventTimestamp": "2024-02-03T14:00:00.000Z",
  "eventType": "transcript",
  "title": "Discovery Call - Acme Corp",
  "transcript": [
    { "speaker": { "name": "Sarah Johnson", "email": "sarah@company.com", "role": "internal" },
      "text": "Thanks for making the time. What pushed you to look at this now?" },
    { "speaker": { "name": "John Smith", "email": "john@acme.com", "role": "external" },
      "text": "Our current process breaks down past about fifty reps." }
  ],
  "participants": [
    { "name": "Sarah Johnson", "email": "sarah@company.com", "role": "internal" },
    { "name": "John Smith", "email": "john@acme.com", "role": "external" }
  ],
  "crmOpportunityId": "opp_12345",
  "eventId": "<stable id derived from my row>"
}

Social event:
{
  "type": "social",
  "eventTimestamp": "2024-02-03T11:00:00.000Z",
  "eventType": "message_sent",
  "body": { "text": "Thanks for connecting, John." },
  "from": { "email": "sarah@company.com", "name": "Sarah Johnson" },
  "to": [{ "email": "john@acme.com", "name": "John Smith" }],
  "eventId": "<stable id derived from my row>"
}

Required fields - everything else is optional:
  email:  type, eventTimestamp, eventType, subject, body, from, to
  call:   type, eventTimestamp, eventType, title, transcript, participants
  crm:    type, eventTimestamp, eventType, opportunityId, opportunityName
  social: type, eventTimestamp, eventType, body, from, to

eventType must be one of:
  email:  sent | reply | opened | clicked | bounced | unsubscribed
  call:   transcript | scheduled | completed | missed
  crm:    opportunity_created | deal_won | deal_lost | meeting_booked
  social: connection_sent | connection_accepted | message_sent | message_received

Only email sent/reply and call transcript are analysed. All CRM and all social
types are analysed. Other email and call types are stored but produce nothing.

## 4. Map my columns onto those structures

Apply these rules:

- eventTimestamp accepts ISO 8601, epoch milliseconds (>= 1e12) or epoch seconds.
  Pass my column straight through if it is already one of those. It must resolve
  within 50 years past and 5 years future - flag sentinel dates (9999-12-31,
  1900-01-01, epoch zero) as rows to drop rather than rows to send.
- Combine separate name and email columns into { "email": ..., "name": ... }.
  Only email is required. Split comma-separated recipient strings into an array.
- Send "amount" as a STRING, not a number, so my decimal precision survives.
- Derive CRM "eventType" from my stage or status column, and set "stageCategory"
  to open, won or lost explicitly rather than letting Octave infer it.
- ONE SOURCE ROW MAY BE SEVERAL EVENTS. A deal row with both a created date and a
  closed date is an opportunity_created event AND a deal_won or deal_lost event,
  each with its own eventTimestamp and its own eventId. Do not collapse them.
- Always send "crmLastModifiedAt" on CRM events. Without it an updated deal can
  hash identical to the previous push and be dropped as a duplicate.
- Prefer a structured transcript array over a flat string. Decide speaker role by
  email domain: my company's domain is "internal", everything else is "external".
- Set "eventId" (<= 255 chars) from the row's stable primary key. Where one row
  expands into several events, suffix it per event so each one stays distinct.
- Populate "contactEmails" with every contact on the deal, not just the primary -
  it is the biggest single lever on how much activity attaches to the opportunity.
  Same for "crmAccountId", "crmAccountDomain" and "lossReason" where I have them.
- On email, call and social rows that already reference a deal, set
  "crmOpportunityId" - it skips Octave's participant-matching inference.
- Omit keys whose source value is NULL. Do not send explicit nulls.

Show me the proposed mapping as a table - my column, Octave field, transformation.
List every Octave field you could not source, and every column of mine you ignored.
Wait for me to confirm the mapping before you write any code.

## 5. Then build and deploy it

Target: POST https://app.octavehq.com/api/v2/event/import
Header: api_key: <my Octave workspace key>
Body:   { "events": [ ... ] }     # max 1000 per request, mixed types allowed

Response: { "jobs": [ { "eventType", "jobOId", "accepted", "preSkipped",
                        "duplicatesInBatch" } ],
            "totalAccepted", "totalPreSkipped" }

One job is created per event type present in the request. Poll each one:
GET https://app.octavehq.com/api/v2/event/import/status?jobOId=<jobOId>
every 5-10 seconds until status is COMPLETED (or FAILED / CANCELLED).
counts.ingested is the authoritative success count. counts.skipped covers
duplicates and non-processable event types and is not an error - alert on
counts.failed instead.

A 400 means NOTHING in that request was imported; the message lists the offending
zero-based indices. Retry 429 and 5xx with exponential backoff, and keep only a
few requests in flight.

Build this for the runtime I named in section 2, and give me:
- the handler code,
- infrastructure-as-code to deploy it - tell me which format you are using and why,
- the watermark held in a managed store rather than on local disk, so a cold start
  resumes where the previous run finished,
- credentials read from the platform's secret manager, never baked into the code,
- structured logs carrying ingested / skipped / failed per batch, and an alert
  condition worth paging on,
- a note on what happens if a run hits the runtime's execution limit, and what to
  change if my history is too large to backfill in a single invocation.

Keep the column mapping in one dict at the top of the handler so I can correct it
without touching the transport or the deployment code.
```

<Tip>
  The instruction to show the mapping as a table and wait for confirmation is the part worth
  keeping. Without it an agent guesses at your column names and buries the guess inside a
  handler, where a wrong `eventType` looks exactly like a right one until the data is in.
</Tip>

## Hints worth adding

These are the things that go wrong with Google Cloud Storage specifically. Paste whichever apply into
section 1 of the prompt — an agent cannot infer them from a schema.

**Parquet timestamps are not milliseconds.** The format stores an explicit unit, and
`us` and `ns` are both common — neither is what Octave reads a large integer as. Ask for
conversion to ISO 8601 at the mapping step.

**Credentials resolve differently in each environment.** Application Default Credentials
mean your account locally and the attached service account in production. Say which
identity the deployed function runs as, or you get something that works in testing and
403s on the first real trigger.

**Eventarc can deliver twice.** Delivery is at-least-once, so the function must be
idempotent. A stable `eventId` already handles it — worth stating so the AI does not
invent a second dedupe layer.