Dependencies & hard gates: sequence the build so it cannot detonate

• Finish-to-start (FS) - B starts only after A finishes. The default and most common: M4 governed DW build starts after M3 legacy migration finishes.
• Start-to-start (SS) - B can start once A has started. Governance definition and infra setup can begin together once the project kicks off; neither waits for the other to finish.
• Finish-to-finish (FF) - B cannot finish until A finishes. Data-quality certification cannot complete until reconciliation completes.
• Mandatory vs discretionary - a hard (mandatory) dependency is physics: you genuinely cannot ingest before the landing zone exists. A soft (discretionary) dependency is preference: you would rather profile all sources before designing, but you could overlap if pushed.

Real dependencies are rarely instantaneous. Lead and lag fine-tune the timing on top of the type.

• Lag - a forced wait after A before B can proceed. A finish-to-start with 2-day lag: reconciliation finishes, then a 2-day sign-off window before certification starts.
• Lead - a permitted head start, letting B begin before A fully finishes. UAT prep can start a few days before migration completes (a negative lag).
• Where it bites: lag hides slack that looks like progress, and lead is where "we overlapped to save time" quietly turns a soft dependency into a risk.

Not every "important" dependency is a gate. A gate passes three tests at once: downstream cannot start until it clears, no amount of overlap buys time, and skipping it forces a rebuild rather than a patch. Northwind's three:

• Hiring gates all technical milestones (M2-M6). No committed technical milestone until the Data Architect and Data Engineering Lead are onboarded. This is the hard dependency named in the charter, now sequenced.
• Governance sign-off gates the M2 architecture. Classification, PII tagging, retention, and access tiers must be signed off before the warehouse design is finalised - or the tiers become a late re-architecture.
• Infra landing zone gates any ingestion. The VPC, IAM, secrets, and security baseline must exist before a single byte lands, or you are ingesting into an unhardened environment.

The most expensive schedule mistake is treating a hard dependency as soft to hit a date. It always looks like progress at first - two workstreams running at once - and the cost lands weeks later when the downstream work has to be redone on the correct foundation.

• The pattern: "We can't wait for governance, let's design the warehouse now and add the tiers after." Then PII surfaces in a "public" schema and the design is re-poured.
• Why it fools people: the parallel start shows visible motion, and the rework cost is deferred past the next status meeting.
• The discipline: before overlapping any link, prove it is soft. If skipping the predecessor would force a rebuild, it is hard - and you wait.

To find the critical path, list the chains of dependent milestones and pick the longest. On Northwind the chain M1 -> M3 -> M4 -> M5 -> M6 is the spine: raw load, then migration, then the governed DW, then exploration, then production. Each finish-to-start link is mandatory, so the chain's total length is the floor on the schedule.

• The rule: the longest dependent chain sets the minimum duration. Adding people to a parallel track does not shorten it - only compressing the critical chain does.
• Slack lives off the path: the hiring, governance, and infra tracks run in parallel early. They have float - but only until they must converge to unblock M2 and the chain.
• Watch the convergence points: a parallel track with slack becomes critical the moment its slack runs out. The hiring track has the least room, because it gates everything technical.

When the critical path is longer than the deadline allows, you have two compression moves - and both trade something.

• Float (slack): the time a non-critical milestone can slip without moving the end date. Total float across a track tells you how much cover you have before it turns critical.
• Fast-tracking: overlapping milestones that were planned in sequence (starting M5 exploration on the certified slice of M4 before all of M4 finishes). It buys time but adds risk, and it is only legal on soft links.
• Crashing: adding resource to a critical milestone to shorten it. Costs money and has diminishing returns - nine engineers do not load an archive in a ninth of the time.