The RSA Role, Decoded

What an RSA actually is

A Resident Solutions Architect is a billable consultant in Databricks Professional Services — not Field Engineering. As of mid-2026, the careers site lists RSA roles under the "professional-services-operations" category. RSAs are embedded with customers on short-to-medium-term post-sales delivery engagements: designing and building reference architectures, productionizing data and AI use cases, and running third-party platform migrations end to end. The job description language is blunt about it — the RSA "completes projects according to specification." This is the SA-family role that actually builds.

Verified from current postings: RSAs scope work alongside Engagement Managers, provide escalated technical support during engagements, coordinate with project managers, other architects, and customer teams, and feed product feedback back to Databricks engineering. Postings are verticalized — Financial Services, Public Sector (US citizenship plus Top Secret clearance eligibility required), Healthcare & Life Sciences, Manufacturing, CME & Games, Digital Native Business — plus Senior RSA ML & GenAI variants that explicitly require RAG, multi-agent systems, Text2SQL, fine-tuning (HuggingFace/LangChain/DSPy), and production GenAI deployment on Mosaic AI. Travel is listed up to roughly 30%. Posted US base salary: $180,656–$248,360 plus bonus and equity; self-reported totals (Glassdoor avg ~$236K, levels.fyi SA-family median ~$363K) suggest meaningful upside, but treat those as directional, not verified.

RSA vs the other "Solutions Architect" titles

Databricks runs four distinct SA-flavored roles, and conflating them in an interview is a credibility hit. The split that matters: presales vs post-sales, and billable vs non-billable.

The week-to-week job

From posting language and candidate/practitioner reports, a typical RSA engagement blends:

• Architecture guidance — designing the target lakehouse, reviewing the customer's existing estate, making build-vs-buy and pattern decisions (verified posting language: "end-to-end design, build, deploy of data and AI applications").
• Hands-on implementation — writing PySpark/SQL, building pipelines, tuning workloads. This is not an oversight role; postings require coding in Python or Scala plus strong SQL and Spark internals depth.
• Escalation handling — postings list "escalated support" during engagements: when the customer's production job falls over mid-project, you are the person in the room.
• Scoping and delivery management — working with Engagement Managers on scope and timeline; "technical project delivery" is an explicit requirement.
• Enablement and trusted-advisor work — commonly reported by practitioners rather than spelled out in postings: upskilling the customer's engineers so the work survives your departure, and earning enough trust to influence roadmap decisions beyond the SOW.
• Utilization targets — because RSAs are billable, your time is tracked against a utilization goal, as in any consulting practice. Specific percentages are commonly reported, not published — ask the hiring manager directly what the target is and how bench time, presales assists, and internal IP work count toward it. Asking shows you understand the PS business model.

The success profile

The 2026 postings ask for 6+ years (Senior ~10+) of data engineering/platforms in customer-facing post-sales or consulting roles; deep Apache Spark including runtime internals; expertise in one cloud and working knowledge of two; CI/CD and MLOps; whiteboarding and documentation; and client management including conflict resolution. Read between the lines and the profile is a three-legged stool:

• Technical depth — Spark execution model, Delta Lake, Unity Catalog, cloud networking and security. Generic warehouse knowledge does not pass; Databricks-specific depth does. (Drill this in Spark internals, Delta Lake, and Unity Catalog.)
• Consulting presence — discovery before solutioning, calm under pushback, comfort being the senior technical voice in a room of strangers, and the ability to say "I don't know, here's how I'd find out" without losing the room.
• Writing and communication — architecture docs, runbooks, SOW-grade scoping language, and status communication that executives can consume. Postings list documentation explicitly; candidate reports say communication is weighted as heavily as code.

Kill any one leg and the stool falls: a brilliant engineer who cannot run a discovery conversation fails the panel; a polished consultant who cannot write live PySpark fails the deep dive.

The interview loop, stage by stage

Candidate reports (Glassdoor, Blind, Taro, Dataford — accessed June 2026) converge on a 5–6 round loop over roughly 4–8 weeks (average ~34 days; one RSA report ran ~2 months). Glassdoor pegs difficulty at 3.4/5 with only ~43% positive experiences — the loop is described as highly selective, with a near-SWE technical bar. Stages vary slightly by report; treat this as the common shape, and confirm specifics with your recruiter.

1. Recruiter screen

Project storytelling — but do not assume it stays soft. At least one report (Taro) describes a recruiter screen diving straight into Spark performance with no warm-up. Prep: have two quantified delivery stories cold (scope, your role, architecture, outcome), and be able to explain a Spark performance win in 90 seconds without notes.

2. Hiring manager

Background depth and customer-facing fit: why Databricks, why Professional Services specifically, how you handle difficult stakeholders. Prep: rehearse the "why RSA, not presales SA or DSA" answer above; bring one story of resolving a client conflict; ask about utilization, vertical alignment, and typical engagement length — it signals you know what PS is.

3. Coding: take-home or CodeSignal (~2–3 hours)

Advanced SQL plus PySpark — complex joins and filtering are the consistent theme; other reports cite hit-counter, game-logic, and IP/CIDR-style problems. Prep: live, timed coding is non-negotiable. Drill window functions and multi-step PySpark transforms until you write them flawlessly without an IDE crutch; candidate reports say needing hints counts against you across the Databricks family of loops.

4. Technical deep dive

Reported coverage: Spark internals (narrow vs wide transformations, Catalyst, memory management, data skew), Delta Lake and ACID, medallion architecture, cloud networking and security, streaming schema evolution, MLOps. Expect scenario framing — "the customer's nightly job suddenly takes 4x longer" — rather than trivia. You should be able to reason about a skewed join out loud and sketch the fix:

# Skewed fact-dim join: diagnose, then pick the cheapest fix that works.
# 1) Confirm skew: a few tasks straggle in the Spark UI; key histogram:
df.groupBy("customer_id").count().orderBy(F.desc("count")).show(10)

# 2) Small dimension? Broadcast it -- no shuffle of the big side:
fact.join(F.broadcast(dim), "customer_id")

# 3) Otherwise let AQE handle it (on by default in modern DBR):
spark.conf.set("spark.sql.adaptive.skewJoin.enabled", "true")

# 4) Last resort for pathological keys: salt the hot keys,
#    explode the dimension across the salt range, join on (key, salt).

Prep: study Databricks-specific tech — Delta internals, Photon, Unity Catalog — not generic warehouse theory, and narrate trade-offs (cost, latency, operational burden), not just mechanisms.

5. Project-delivery / consulting discussion

Scoping, timeline management, what you do when scope slips or the customer's team underdelivers. Prep: structure answers as situation, options considered, decision, outcome — and always name the trade-off you accepted.

6. Panel presentation / customer simulation

The decisive round — covered in full below. Some loops add an executive round and a hiring-manager wrap-up after it.

Inside the panel presentation

Consistently cited as the hardest round. You get a business scenario in advance, build a recommendation, and present to a panel roleplaying customer stakeholders who push back deliberately on cost and complexity. Reports say it evaluates four things:

• Discovery — do you ask qualifying questions before solutioning, or dump an architecture on minute one? Open by confirming business goals, current pain, constraints, and success criteria.
• Structure — a clear arc: problem, options, recommendation, phased plan, risks. Rehearse the slide and screen-share mechanics; fumbling logistics burns credibility minutes.
• Trade-off reasoning — including competitive positioning: why Databricks over Snowflake, Redshift, or DIY Spark for this workload, framed in business value and cost, with honest acknowledgment of where alternatives are fine.
• Handling pushback — "this looks expensive," "our team can't operate this," "why not just keep the warehouse?" The test is composure: acknowledge, quantify, offer a phased de-risked path. Never get defensive, never bluff.

Mapping your background to the role

The RSA posting reads like a checklist of what you already do at ADM — minus the consulting business card. The mapping you should make explicit in every round:

• Embedded delivery ownership — you lead Finance R2R analytics delivery end to end: architecture through hypercare. That is the RSA delivery lifecycle, with ADM as the "customer."
• Stakeholder discovery and fit-gap — you run discovery with finance stakeholders and produce fit-gap and architecture recommendations (batch vs near-real-time, CDC/SCD design, Power BI vs Dremio/Denodo semantic layers). That is consulting work in everything but title.
• Runbooks, hypercare, enablement — CI/CD, monitoring, runbooks, and hypercare through Azure DevOps; mentoring junior engineers through design and code reviews. Enablement-so-it-survives-your-departure is the trusted-advisor muscle.
• Production GenAI — three shipped agents on Databricks over HR data: a Genie Space headcount/salaries agent with the SQL shown for trust, a policy RAG assistant on Vector Search with cited answers, and a multi-agent router, all delivered as a Databricks App with Unity Catalog row- and column-level security. That maps directly onto the Senior RSA ML & GenAI requirements (RAG, multi-agent, Text2SQL, production deployment).
• Migrations — Maersk SSAS-to-Azure Analysis Services, Port Authority VBA/Access-to-Synapse, Microsoft on-prem SSIS-to-Azure: three legacy-to-cloud migrations across three employers, which is the bread and butter of PS engagements (see Migration Playbooks).

The honest gap analysis

Interviewers respect a candidate who names their own gaps before being asked. Yours, and how to handle each:

One closing calibration: the loop is long, the bar is high, and reported pass sentiment is under 50% — so treat every round, including the recruiter screen, as technical until proven otherwise, and treat the panel as the round you over-prepare. The role itself is the closest thing Databricks has to what you already do all day; the interview is about proving you can do it in someone else's building, on the clock, with the meter running.