system-type-multi-tenant-saas

System Type: Multi-Tenant SaaS

Patterns, failure modes, and anti-patterns for multi-tenant software-as-a-service platforms.

---

1. Isolation Models

Every multi-tenant system lives on a spectrum between full sharing and full isolation. The choice is not one-time — most mature SaaS platforms mix models across their stack.

The Spectrum

When to Mix Models

The right answer is almost always a hybrid. Small tenants on shared infrastructure, enterprise tenants on dedicated. The boundary is driven by:

• Revenue: A tenant paying $50/month cannot justify a dedicated database. A tenant paying $500K/year can justify a dedicated cluster.
• Compliance: HIPAA, SOC2 Type II, FedRAMP, and data residency laws (GDPR Article 44+, China's PIPL) may require physical isolation regardless of cost.
• Risk tolerance: If a single tenant's data breach would be existential (healthcare, finance), isolate by default.
• Noisy neighbor exposure: If one tenant can generate 100x the load of a median tenant, shared compute is a liability.

The Compliance Driver

Compliance is the most common reason teams are forced up the isolation spectrum against their economic interest. Understand these constraints early:

• Data residency: Some tenants require data to stay in a specific geographic region. This forces at minimum a separate data store, often a separate deployment.
• SOC2 / ISO 27001: Auditors want to see logical or physical isolation. Row-level tenant_id filtering is defensible but requires strong evidence (query audit logs, automated testing, access controls).
• HIPAA: Business Associate Agreements (BAAs) often require encryption at rest per tenant and audit trails. Shared-everything makes this operationally painful.
• FedRAMP: Usually requires a fully separate environment. Not a spectrum decision — it's a separate product deployment.

Practical Guidance

Design for shared-everything first, but make the isolation boundary a first-class abstraction. Every data access path should go through a tenant context that can be swapped from "row filter" to "connection router" without rewriting application code. Teams that don't do this spend 6-18 months retrofitting isolation when their first enterprise customer demands it.

---

2. Data Partitioning

Data partitioning is where multi-tenancy gets real. The schema design you pick on day one will constrain your options for years.

Row-Level Isolation (tenant_id Columns)

The most common starting point. Every table has a tenant_id column. Every query includes a WHERE tenant_id = ? predicate.

What goes wrong:

• Missed predicates: One query without a tenant_id filter leaks data across every tenant. This is the single most common multi-tenant security vulnerability. Mitigate with: PostgreSQL Row-Level Security (RLS), ORM-level query interceptors that inject tenant filters, automated test suites that assert every query plan includes the tenant predicate.
• Index bloat: Without composite indexes that lead with tenant_id, queries scan far more data than needed. Every index should be (tenant_id, ...) not (..., tenant_id).
• Hot partitions: If you use database-level partitioning by tenant_id, large tenants create hot partitions. Consider range or hash partitioning across tenant groups instead.
• Backups and restores: You cannot restore a single tenant's data from a shared database backup without extracting their rows. This makes tenant-level disaster recovery painful.

Schema-Per-Tenant

Each tenant gets their own schema (e.g., PostgreSQL schemas) within a shared database instance. Application code sets search_path per request.

Tradeoffs:

• Stronger isolation than row-level. A bug in tenant A's query cannot accidentally touch tenant B's tables.
• Schema migrations must be applied N times. A migration that takes 2 seconds per schema takes 2 hours with 3,600 tenants. Migrations must be non-blocking and idempotent.
• Connection pooling becomes complex. PgBouncer's search_path switching is fragile. Consider one pool per schema or a connection-per-request model with SET statements.
• Monitoring blind spots: pg_stat_statements aggregates across schemas. You need per-schema query metrics to identify slow tenants.

Database-Per-Tenant

Each tenant gets a dedicated database instance (or cluster).

Tradeoffs:

• Maximum isolation. Independent backup/restore, independent scaling, independent maintenance windows.
• Connection management at scale is the primary challenge. 1,000 tenants means 1,000 database instances. Use a connection routing layer (e.g., ProxySQL, PgBouncer with routing rules, or application-level connection pools keyed by tenant).
• Schema migrations require orchestration across all instances. A failed migration on instance 437 of 1,000 must not block the other 999. Use a migration state machine that tracks per-instance progress.
• Cost is significant. Even with small instances, the operational overhead of monitoring, patching, and managing thousands of databases is real.

Partition Key Selection

tenant_id is the obvious partition key, but the right physical partitioning depends on access patterns:

• If most queries are tenant-scoped (the common case), partition by tenant_id. This gives you data locality and the option to move a tenant's partition to dedicated storage.
• If you have significant cross-tenant queries (analytics, admin dashboards), partitioning by tenant_id makes those queries expensive — they must scan every partition. Maintain a separate analytical store for cross-tenant access.
• Composite partition keys (tenant_id + time range) work well for time-series data within tenants. Partition by tenant for isolation, sub-partition by time for retention and query performance.

Cross-Tenant Query Prevention

Regardless of partitioning model, enforce tenant boundaries at multiple layers:

1. Application layer: Tenant context injected at request entry, propagated through the call stack. Never allow tenant ID to come from user input in a data query — resolve it from the authenticated session.
2. ORM / data access layer: Global query filters or scopes that automatically inject tenant predicates. In Django, use a custom manager. In SQLAlchemy, use events or a session-level filter. In Rails, use default_scope cautiously.
3. Database layer: PostgreSQL RLS policies, MySQL views with SECURITY DEFINER, or equivalent. This is your safety net when application-layer filtering has a bug.
4. Audit layer: Log every query's tenant context. Run periodic reconciliation that scans for queries missing tenant predicates.

Migrating Between Isolation Levels

The most painful operation in a multi-tenant system. Typical path: row-level → schema-per-tenant → database-per-tenant as a tenant grows.

Requirements for making this possible:

• Tenant routing must be a configuration lookup, not a code path. The application asks "where is tenant X's data?" and gets a connection config back.
• Data migration must be online. Double-write during migration: write to both old and new location, backfill historical data, switch reads, stop writes to old location, clean up.
• Always have a rollback plan. The new isolation level may perform worse for that tenant's access patterns.

---

3. Noisy Neighbor Mitigation

In shared infrastructure, one tenant's workload affects every other tenant. This is the defining operational challenge of multi-tenancy.

Per-Tenant Rate Limiting

Rate limiting is the first defense. Apply at multiple granularities:

• API rate limits: Requests per second/minute per tenant. Use token bucket or sliding window. Leaky bucket is too permissive for burst-heavy tenants.
• Compute rate limits: CPU-seconds or request-duration budgets per tenant per time window. A tenant running expensive reports should not starve tenants running simple CRUD.
• Storage rate limits: Write throughput, storage growth rate. A tenant bulk-importing data can saturate disk I/O for the entire instance.
• Background job limits: Concurrent job slots or queue priority per tenant. Without this, one tenant's bulk export blocks another's critical webhook delivery.

Resource Quotas

Rate limits control flow. Quotas control total allocation:

• Maximum rows per table, maximum storage per tenant, maximum concurrent connections.
• Quotas should be soft by default (warn, then throttle) and hard only for system-protection scenarios.
• Always expose quota usage to tenants via API and dashboard. Surprise quota enforcement destroys trust.

Fair Scheduling

For shared compute resources (worker pools, job queues, report generators):

• Weighted fair queuing: Each tenant gets a share of processing capacity proportional to their tier. A free-tier tenant gets 1 share, an enterprise tenant gets 10 shares.
• Per-tenant queues with round-robin dispatch: Each tenant's work goes into a dedicated queue. Workers pull from queues in round-robin order. This prevents one tenant's 10,000-item batch from blocking another's single item.
• Preemption for latency-sensitive work: Interactive requests preempt batch work. Tag requests with priority classes.

Detecting Noisy Neighbors

You cannot mitigate what you cannot see. Instrument:

• Per-tenant CPU time, memory allocation, I/O operations, query count, and query latency at the application tier.
• Per-tenant database metrics: rows scanned, temp table usage, lock wait time, connection count.
• Per-tenant queue depth and processing latency.

Set anomaly detection thresholds: alert when a tenant's resource consumption exceeds 3x their historical P95 or when their usage exceeds 50% of total shared capacity.

Throttling vs Queuing vs Rejection

Always prefer throttle → queue → reject in that order. Rejection should only happen when platform stability is at risk.

Tier-Based Resource Allocation

Different pricing tiers get different resource envelopes:

• Define resource profiles per tier: API rate limits, compute budgets, storage quotas, job concurrency.
• Enforce at the gateway layer (API rate limits) and the application layer (compute budgets, job concurrency).
• Make the limits visible in the pricing page. Hidden limits discovered in production create adversarial customer relationships.

---

4. Per-Tenant Configuration

Customization is what makes SaaS sticky. It's also what makes SaaS unmaintainable.

Configuration Inheritance

Use a layered configuration model with clear precedence:

Platform defaults → Plan/tier defaults → Tenant overrides → User preferences

Each layer only specifies deltas from the layer below. Resolution logic:

1. Start with platform defaults (every tenant gets these).
2. Overlay plan-level defaults (e.g., "Enterprise" plan enables SSO by default).
3. Overlay tenant-specific overrides (e.g., tenant X has a custom session timeout).
4. Overlay user-level preferences where applicable (e.g., notification settings).

Store configuration as a typed document per tenant, not as rows in a key-value table. Key-value configuration tables always devolve into stringly-typed chaos that no one can reason about.

Feature Flags Per Tenant

Feature flags in multi-tenant systems serve two purposes:

1. Rollout gating: Gradually enable a feature across tenants (canary → 10% → 50% → 100%).
2. Plan entitlements: Feature X is available on Professional and above.

Keep these separate. Rollout flags are temporary and should be cleaned up. Entitlement flags are permanent and are part of your billing model.

Implementation:

• Evaluate flags at request time, not build time. The tenant context determines flag state.
• Cache flag evaluations per tenant with a short TTL (30-60 seconds). Don't hit the flag store on every request.
• Maintain a flag manifest that lists every flag, its type (rollout vs entitlement), its owner, and its expected cleanup date. Flags without owners accumulate forever.

Custom Branding and Configurable Workflows

These are the features sales teams promise and engineering teams regret:

• Custom branding (logos, colors, email templates): Build a theming system with a defined schema. Allow CSS variable overrides, not arbitrary CSS injection. Arbitrary CSS is a security risk (CSS injection can exfiltrate data) and a maintenance burden.
• Custom workflows (approval chains, notification rules, field mappings): Build a workflow engine with composable primitives. Do not allow arbitrary code execution per tenant. Every "custom workflow" system eventually becomes a bad programming language. Constrain the primitives.
• Custom fields / schema extensions: Allow tenants to add custom fields to core entities. Store as a typed JSON document with a schema, not as EAV (Entity-Attribute-Value) tables. EAV tables destroy query performance and make reporting impossible.

The Complexity Explosion

Every per-tenant configuration option multiplies your test matrix. 10 boolean flags = 1,024 combinations. You cannot test them all.

Mitigations:

• Define supported configuration profiles (e.g., "Standard," "Healthcare," "Financial Services") and test those combinations exhaustively. Allow custom overrides but disclaim that novel combinations are best-effort.
• Build a configuration linter that detects incompatible or nonsensical combinations (e.g., SSO enabled but no identity provider configured).
• Log the active configuration at request time so that bug reports can be correlated with tenant configuration state.

---

5. Billing and Metering

Billing is where abstract architecture meets concrete revenue. Errors here cost real money and erode trust.

Usage Metering Architecture

Usage metering has a deceptively simple requirement: count how much each tenant used, accurately, and turn that into a bill. In practice:

Event ingestion → Deduplication → Aggregation → Rating → Invoice generation

Event ingestion: Every billable action emits a metering event. Events must include: tenant ID, event type, quantity, timestamp, and an idempotency key.

Deduplication: Events will be delivered more than once (retries, replays, infrastructure failures). Use the idempotency key to deduplicate. Store processed event IDs for at least one billing cycle.

Aggregation: Sum events per tenant per billing dimension per billing period. Use tumbling or hopping windows depending on your billing granularity.

Rating: Apply the tenant's pricing plan to the aggregated usage. Handle mid-cycle plan changes (prorate, or apply new plan from next cycle — decide this in advance and document it).

Invoice generation: Generate invoices at billing cycle boundaries. Handle edge cases: tenant created mid-cycle, plan changed mid-cycle, metering data delayed past cycle close.

Idempotent Meter Events

This is the single most important property of your metering system. If you cannot safely replay metering events without double-counting, your billing is wrong.

Requirements:

• Every event has a globally unique idempotency key, generated at the source.
• The metering pipeline deduplicates on this key before aggregation.
• Deduplication window must exceed the maximum expected replay window (usually 24-72 hours).
• Test this explicitly: replay a day's events and verify zero billing change.

Billing Models

Most SaaS products start seat-based for simplicity and migrate to hybrid as they mature. Pure usage-based pricing requires excellent cost attribution and real-time usage dashboards to avoid bill shock.

Aggregation Windows and Edge Cases

• Clock skew: Events from different services may have slightly different timestamps. Use server-side receive time for billing, not client-reported event time. Allow a grace period (e.g., 5 minutes) at window boundaries.
• Late-arriving events: An event for March arrives on April 2nd. Policy decision: include in March's bill (requires reopening the billing cycle) or include in April's bill (simpler but less accurate). Document and be consistent.
• Billing cycle boundaries: If a tenant's cycle closes at midnight UTC and a request spans 23:59:58 to 00:00:03, which cycle does it belong to? Use the start time of the request.
• Free tier and trial: Meter everything, even if the price is zero. You need the data for capacity planning and conversion analysis. Never skip metering for free-tier tenants.

Revenue Recognition

For SaaS with annual contracts or prepaid credits, revenue recognition (ASC 606 / IFRS 15) adds constraints:

• Recognize revenue as the service is delivered, not when payment is received.
• Prepaid credits that expire create deferred revenue liabilities.
• Usage-based revenue is recognized in the period the usage occurs.
• Finance will need metering data sliced by recognition period. Design the data model to support this from the start, or face a painful retrofit at Series B when auditors ask for it.

---

6. Tenant Lifecycle

A tenant is not just a row in a table. It's a lifecycle with provisioning, scaling, migration, and eventual termination — each phase with its own failure modes.

Onboarding Automation

Tenant creation should be fully automated and idempotent. The provisioning pipeline:

1. Create tenant record in the control plane database. Assign a globally unique tenant ID (UUID, not auto-increment — you'll merge databases someday).
2. Provision data storage: Create schema, database, or initialize row-level isolation structures.
3. Provision infrastructure (if siloed): Spin up dedicated compute, configure networking, set up DNS.
4. Initialize configuration: Apply plan defaults, set up default admin user, generate API keys.
5. Seed reference data: Default categories, templates, sample data if applicable.
6. Verify: Run a health check against the new tenant's environment. Automated smoke test.

Each step must be idempotent. If step 4 fails and you retry, steps 1-3 must not create duplicates. Use a state machine with explicit states: PENDING → PROVISIONING → ACTIVE → SUSPENDED → TERMINATED.

Tenant Provisioning Failure Handling

Provisioning will fail. Database creation times out, infrastructure quota exceeded, DNS propagation delayed. Handle with:

• A provisioning state machine that records which steps completed successfully.
• Automatic retry with exponential backoff for transient failures.
• A cleanup path that tears down partial provisioning (orphaned databases, dangling DNS records).
• Alerting on provisioning failures — a failed tenant creation is a lost customer.

Data Migration Between Tiers

When a tenant upgrades from shared to dedicated infrastructure:

1. Create the target infrastructure.
2. Begin dual-writing: all new writes go to both old and new locations.
3. Backfill historical data from old to new location. Use checksums to verify completeness.
4. Verify data consistency between old and new locations.
5. Switch reads to the new location (feature flag per tenant).
6. Stop writes to the old location.
7. Retain old data for a defined period (30-90 days), then clean up.

This is a multi-day operation for large tenants. Build it as an automated pipeline with human approval gates at steps 5 and 7.

Tenant Offboarding and Data Deletion

Offboarding is the most legally consequential lifecycle operation.

GDPR Right to Erasure (Article 17) / CCPA / other data protection laws:

• You must be able to delete all of a tenant's data upon request. "All" includes: primary data store, backups, logs, analytics pipelines, third-party integrations, CDN caches, search indexes, message queues, data warehouses.
• Backups are the hardest part. You cannot surgically remove one tenant's data from a shared backup. Options: accept that backup data will be retained until the backup expires (document this in your DPA), or use tenant-level encryption with key destruction (delete the key, the data becomes unreadable).
• Log retention: Tenant data in logs must be either deletable or anonymized. This requires structured logging with tenant-scoped retention policies.
• Maintain a deletion audit trail: record what was deleted, when, and from which systems. The audit trail itself must not contain the deleted data.

Practical offboarding pipeline:

1. Suspend tenant (block all access, stop billing).
2. Generate data export (see below).
3. Wait for export download confirmation or expiry period.
4. Execute deletion across all systems. Track completion per system.
5. Verify deletion. Run queries against all data stores to confirm zero rows for the tenant.
6. Retain the deletion audit record.

Tenant Data Export

Tenants will want their data out. Regulatory requirements (GDPR data portability) may mandate it.

• Provide a bulk export in a standard format (JSON, CSV). Not your internal database dump — a documented, versioned export schema.
• Large exports should be asynchronous: tenant requests export → system generates → tenant receives download link.
• Include all tenant data: primary entities, configuration, audit logs, uploaded files.
• Rate-limit export generation to prevent abuse (export is an expensive operation).

---

7. Observability in Multi-Tenant Systems

Standard observability breaks down in multi-tenant systems. Platform-wide metrics hide tenant-specific problems. Tenant-specific debugging risks exposing another tenant's data.

Tenant-Aware Logging

Every log line must include the tenant ID. This is non-negotiable. Without it, debugging tenant-specific issues requires correlating timestamps and hoping.

Implementation:

• Inject tenant ID into the logging context at request entry (MDC in Java, contextvars in Python, AsyncLocalStorage in Node.js).
• Propagate tenant ID through async boundaries (message queues, background jobs, scheduled tasks).
• Structured logging (JSON) with tenant_id as a top-level field. Not buried in the message string.

Access control on logs: Engineers debugging tenant A's issue should not see tenant B's log data. Use log partitioning by tenant or field-level access controls in your log aggregator. In practice, most teams settle for "logs are accessible to on-call engineers under audit" — but document this in your security posture.

Per-Tenant Metrics

Emit metrics with a tenant_id label/tag. But be aware of cardinality:

• 10,000 tenants × 50 metrics × 5 label dimensions = cardinality explosion. Prometheus will fall over. Datadog will send you a bill that looks like a phone number.
• Solution: Emit high-cardinality per-tenant metrics to a separate pipeline (ClickHouse, BigQuery) for analysis. Keep low-cardinality aggregate metrics in your real-time monitoring system.
• Pre-aggregate per-tenant metrics at the application tier: compute P50/P95/P99 latency per tenant in-process and emit the summary, not every individual observation.

SLA Tracking Per Tier

Different tiers have different SLA commitments. Track compliance per tier:

• Define SLIs (Service Level Indicators) per tier: availability, latency P99, error rate.
• Compute SLOs (Service Level Objectives) per tier: "Enterprise tier: 99.95% availability, P99 latency < 200ms."
• Track error budgets per tier. If the enterprise tier has consumed 80% of its monthly error budget, that's a higher priority than a free-tier outage.
• Report SLA compliance per tenant to customer success teams. Don't wait for the customer to complain.

Debugging Tenant-Specific Issues

The hardest observability problem: tenant X reports slowness, but your platform-wide dashboards look fine.

Approach:

1. Filter all metrics and logs to the reporting tenant. Look for elevated latency, error rates, or resource consumption.
2. Compare the tenant's current behavior to their historical baseline. Are they doing something different (more users, larger payloads, new integration)?
3. Check for noisy neighbors: is another tenant on the same infrastructure consuming disproportionate resources?
4. Check tenant-specific configuration: did a recent config change (feature flag, custom workflow) alter behavior?
5. Reproduce in an isolated environment if possible. Never reproduce in production with another tenant's data.

Build a tenant health dashboard that on-call engineers can pull up immediately: current request rate, error rate, latency, active users, recent configuration changes, infrastructure placement (which shard, which cluster).

---

8. Common Failure Modes

These are the failures every multi-tenant SaaS platform encounters. Plan for them.

Cross-Tenant Data Leakage

Cause: Missing or incorrect tenant filtering in a query, API response including data from another tenant, cache key collision across tenants, search index returning results from wrong tenant.

Impact: Existential. One confirmed cross-tenant data leak can trigger customer exodus, regulatory action, and breach notification requirements.

Prevention: Defense in depth — application-layer filtering, database-layer RLS, automated testing (create data as tenant A, attempt to read as tenant B, assert zero results), periodic cross-tenant audit scans, cache keys always prefixed with tenant ID.

Noisy Neighbor Cascading Failure

Cause: One tenant's workload (large import, expensive report, runaway integration) consumes shared resources. Database connection pool exhausted, CPU saturated, memory pressure triggers OOM kills.

Impact: Every tenant on the shared infrastructure experiences degraded performance or outages.

Mitigation: Per-tenant resource limits (connection pool partitioning, CPU cgroups, memory limits), circuit breakers per tenant, automatic load shedding that sacrifices the noisy tenant before the platform. Kill switches: the ability to instantly disable a specific tenant's access.

Tenant Migration Data Loss

Cause: During migration between isolation levels, data written to the old location after the cutover is lost. Backfill is incomplete. Referential integrity broken by partial migration.

Impact: Tenant data loss or corruption, potentially undetectable until the tenant reports it.

Mitigation: Dual-write during migration, checksums on backfill, reconciliation reports before cutover, rollback capability for at least 30 days post-migration.

Billing Metering Gaps

Cause: Metering events dropped during infrastructure failures, clock skew causes events to fall outside aggregation windows, deduplication logic incorrectly discards valid events, new features shipped without metering instrumentation.

Impact: Revenue leakage (under-billing) or customer disputes (over-billing). Under-billing is more dangerous because it's invisible until you audit.

Mitigation: Metering completeness checks (compare expected vs actual event counts), alert on metering pipeline lag, require metering instrumentation as part of the feature launch checklist, periodic billing reconciliation against actual resource consumption.

Configuration Drift Between Tenants

Cause: Manual configuration overrides accumulate over time. Tenant A was given a special setting in 2022 that no one remembers. A platform-wide default change doesn't apply to tenants with overrides.

Impact: Inconsistent behavior across tenants. Bugs that only reproduce for specific tenants. Support burden increases as every ticket requires configuration investigation.

Mitigation: Configuration audit trail (who changed what, when, why), configuration drift detection (alert when a tenant's config diverges from plan defaults by more than N overrides), periodic configuration review with customer success.

Per-Tenant Schema Migration Failures

Cause: In schema-per-tenant or database-per-tenant models, a migration succeeds for 4,997 tenants and fails for 3. The failures are due to data-dependent edge cases (constraint violation on unexpected data, timeout on large tables).

Impact: Tenants on different schema versions. Application code that assumes the new schema breaks for the failed tenants. Rolling back the migration for all tenants is often impractical.

Mitigation: Migrations must be backward-compatible (add columns, don't remove/rename). Track migration state per tenant independently. Build a migration retry and remediation pipeline. Test migrations against a copy of production data for the largest tenants before deploying.

---

9. Anti-Patterns

"We'll add tenant_id later." Retrofitting tenant isolation into an application built without it is a 6-12 month project that touches every query, every cache key, every background job, and every test. The tenant_id must be in the data model from day one, even if you only have one tenant.

"Every tenant gets their own deployment." Sounds like maximum isolation. In reality, it means N deployments to update, N sets of infrastructure to monitor, N configurations that drift. This is not multi-tenancy — it's managed hosting. It works at 5 tenants. It collapses at 50. If you must do it, invest in deployment automation that treats tenants as cattle, not pets.

"The ORM handles tenant filtering." ORMs are leaky abstractions. A raw SQL query, a bulk operation, a migration script, a reporting query — any code path that bypasses the ORM bypasses your tenant filtering. Tenant isolation must be enforced at the database layer (RLS, views) as a backstop, not solely at the application layer.

"We'll just use a separate schema for everyone." Schema-per-tenant works up to low hundreds of tenants. Beyond that, connection management becomes a nightmare, schema migrations take hours, and database tooling (monitoring, backups, query analysis) breaks down. Know your ceiling before you commit.

"Per-tenant customization is just configuration." It starts with "let them change the logo." Then it's "let them customize the email template." Then it's "let them define custom approval workflows." Then it's "let them write custom validation rules." Each step feels incremental. The cumulative result is a platform where no two tenants behave the same way, bugs are unreproducible, and every change requires testing against an exponential configuration matrix. Set hard boundaries on customization scope. Build a platform, not a programming environment.

"Usage metering can be eventually consistent." It can, but "eventually" must be bounded and well-understood. If your metering pipeline has a 4-hour lag and a tenant's billing cycle closes during that lag, you either under-bill or reopen the cycle. Neither is good. Define the maximum acceptable metering delay and engineer the pipeline to meet it with margin.

"Free-tier tenants don't need metering." They do. Free-tier tenants are your biggest source of noisy neighbors (no payment friction means no natural usage constraint). Meter them, quota them, and rate-limit them more aggressively than paid tenants, not less.

"We can just query across tenants for analytics." Cross-tenant analytics queries on your production OLTP database will bring it to its knees. Replicate to an analytical store (data warehouse, read replica) and run cross-tenant queries there. Better yet, pre-aggregate cross-tenant metrics in a pipeline so no single query needs to scan all tenants.

"Tenant deletion is just a soft delete." Soft delete is not deletion under GDPR. A deleted_at timestamp does not satisfy the right to erasure. You must actually remove the data from primary stores, and have a plan for backups, logs, and downstream systems. Soft delete is a useful intermediate step (suspend first, delete later), but it is not the end state.

"One connection pool is fine." A single shared connection pool means one tenant running 50 concurrent slow queries can exhaust connections for every other tenant. Partition connection pools by tenant or by tenant tier. At minimum, reserve a pool of connections that no single tenant can fully consume.