Flagged cards

Backend & Data Engineering basic

Front What are advisory locks for, when row locks already exist? Back Locking things that aren't rows: "only one instance of this cron job", "one migration at a time", "one rebuild per tenant". The lock key is just a number you invent — the database enforces mutual exclusion without needing a table to hang it on. pg_try_advisory_lock returning false is the polite version of "another instance is already running, exit 0".
Business & Marketing basic

Front Plenty of early-stage opportunities, very few reaching late stage. First hypothesis? Back Discovery is passing unqualified deals forward — stage 2 is being awarded for interest rather than for a compelling event and an identified buyer. The fix is at the entry criteria, not at the stage where the deals visibly die.
Backend & Data Engineering basic

Front Queue depth is growing, but the consumers are mostly idle. First hypothesis? Back The work is unbalanced or stuck, not excessive. Suspects: a hot partition key sending everything to one consumer while its siblings starve; a poison message being redelivered in a loop at the head; or a visibility-timeout/prefetch misconfiguration parking messages invisibly. Idle consumers acquit throughput and indict distribution — scaling out consumers, the reflex fix, changes nothing.
Business & Marketing basic

Front MRR growing, gross margin down 4 points over two quarters. First hypothesis? Back Mix shift toward larger customers carrying heavier support and infrastructure cost, or discounting to close them. Check margin per segment; blended margin hides it by construction.
Business & Marketing basic

Front Roughly what share of a 10-year DCF's value sits in the terminal value? Back Sixty to eighty percent. The input you are least confident about dominates the answer, which is why DCF sensitivity tables matter more than the point estimate.
Backend & Data Engineering basic

Front What does a UUIDv4 primary key cost on a high-insert table, and what fixes it while keeping UUIDs? Back Random keys insert into random B-tree pages: every insert touches a cold page, page splits multiply, cache hit rate collapses, and WAL volume grows with full-page writes. Sequential keys append to the same hot rightmost page. UUIDv7 (or ULID) fixes it: a timestamp prefix makes keys roughly ordered, restoring append-like locality while keeping global uniqueness and non-guessability-in-practice. Same 16 bytes, most of the benefit of a bigserial.
Web Performance & UX basic

Front CLS is fine on desktop but bad on mobile. What shifts things at narrow widths? Back The usual culprits are elements whose height is only unknown on small screens: images without width/height attributes (the wide layout hid it, the single column doesn't), injected banners and ads whose slots have no reserved height, and web fonts swapping in at a different metric so text reflows — all magnified because on mobile every element is above the fold. The unifying fix: reserve the space before the content arrives — aspect-ratio boxes for media, min-height for ad slots and banners, size-adjusted font fallbacks. CLS is almost never a rendering bug; it's a promise about layout you didn't make.
Business & Marketing basic

Front SaaS quick ratio — the formula and the threshold? Back (New MRR + expansion MRR) ÷ (churned MRR + contraction MRR). Above 4 means you are adding four euros of growth for every one you lose. It is the cleanest single read on whether growth is being manufactured or merely replacing losses.
Backend & Data Engineering basic

Front What anomaly does REPEATABLE READ still permit that SERIALIZABLE prevents — and what's the canonical example? Back Write skew: two transactions read an overlapping set, each writes to a *different* row, and the combined result violates an invariant neither could see breaking alone. On-call doctors: rule says ≥1 must be on call. Both are. Each doctor's transaction checks "how many on call?" → sees 2 → removes *themselves*. Neither wrote to a row the other read, so REPEATABLE READ finds no conflict. Now zero doctors are on call.
Backend & Data Engineering basic

Front What is a retry budget, and what failure does it prevent that per-request retry caps don't? Back A cap on retries as a *fraction of total traffic* (say 10%) — beyond it, failures return immediately. Per-request caps ("3 tries each") still let system-wide load triple exactly when the downstream is drowning, because every request maxes its individual allowance at once. The principle: retries are a tax the healthy pay to smooth blips. When failure is systemic, the tax becomes the attack; the budget detects "everyone is retrying" and stops.
Backend & Data Engineering basic

Front Which HTTP errors should never be retried, and which is the special case? Back Don't retry what will fail identically: 400, 401, 403, 404, 422 — the request itself is wrong, and retrying is load without hope. Retry 5xx and network timeouts. The special case is 429: retry, but on the server's schedule (Retry-After), not yours — it's the one error that is explicitly *about* your retry behaviour. The sharp edge: a timeout on a non-idempotent POST is not safely retryable at all until an idempotency key makes it so — the request may have succeeded.
Backend & Data Engineering basic

Front An idempotent consumer via processed-events table — what makes the pattern crash-proof? Back Insert the event id into a processed table (unique constraint) *in the same transaction* as the effect. Duplicate arrives → constraint fires → skip. Crash after commit → redelivery hits the constraint. Crash before commit → both the effect and the marker roll back, so redelivery correctly re-runs. The atomicity is the whole pattern: mark-then-process loses events on crash, process-then-mark duplicates them. Same transaction, or it doesn't work.
Business & Marketing basic

Front Little's Law, applied to a sales pipeline? Back Open deals = deal arrival rate × average cycle time. So a pipeline of 60 opportunities with a 3-month cycle can only produce ~20 closes a month at 100% win rate — no amount of pipeline review changes that arithmetic without changing arrival rate or cycle time.
Business & Marketing basic

Front Sample ratio mismatch — what is it and why does it invalidate a test? Back Your 50/50 split delivered 52/48. With large n that is statistically impossible by chance, so something in assignment, logging or redirects is broken — and whatever broke is almost certainly correlated with the outcome. An SRM check is the first thing to run on any surprising result.
Business & Marketing basic

Front The general sample-size rule of thumb for a two-arm test? Back n per arm ≈ 16σ² ÷ Δ², where Δ is the absolute effect you want to detect. For a proportion, σ² = p(1−p). The 16 encodes 80% power at 95% confidence; it is not a magic constant, it is (1.96 + 0.84)² × 2 rounded.
Web Performance & UX basic

Front The Core Web Vitals thresholds — good and poor, for all three (as of 2025)? Back LCP: good ≤ 2.5s, poor > 4s. CLS: good ≤ 0.1, poor > 0.25. INP: good ≤ 200ms, poor > 500ms. Between the bounds is "needs improvement". The part everyone forgets: the assessment is at the *75th percentile of field data* — your page passes only if three-quarters of real visits pass. A fast median with a bad tail fails, which is deliberate: the tail is where slow devices and networks live, and that's most of the mobile world.
Business & Marketing basic

Front The winner's curse in A/B testing — what is it and what should you expect? Back The variant that wins a test is the one whose noise happened to point up, so the measured lift is biased high. Expect roughly half the observed effect to survive into production — and budget the rollout on that, not on the test readout.
Business & Marketing basic

Front You want to detect a 10% relative lift on a 5% baseline conversion rate. How much traffic per variant? Back Roughly 30,000 per arm, for 80% power at 95% confidence. n ≈ 16 × p(1−p) ÷ Δ² = 16 × 0.0475 ÷ 0.005² ≈ 30,400. This is the most useful number in CRO, because for most companies it is an honest answer of "you cannot A/B test your way to small wins".
Backend & Data Engineering basic

Front How does Postgres SERIALIZABLE detect conflicts without locking everything? Back It tracks what each transaction *read* (predicate-level SIRead locks, which block nobody) and looks for cycles of read-write dependencies between concurrent transactions. When it finds a structure that could produce a non-serializable outcome, it aborts one participant. Detection is conservative: false positives happen, which is another reason the retry loop is mandatory rather than optional.
Backend & Data Engineering basic

Front What happens if you ask Postgres for READ UNCOMMITTED? Back You silently get READ COMMITTED. MVCC never exposes uncommitted data to anyone — dirty reads are structurally impossible in Postgres, so the weakest standard level has nothing weaker to offer.
Business & Marketing basic

Front Rule of 40 — the formula, precisely? Back Year-over-year revenue growth % plus free cash flow margin % should be at least 40. EBITDA margin is the common substitute; state which one you are using, because they can differ by 10+ points.
Business & Marketing basic

Front Two products, identical month-1 retention. Which is more valuable, and what is the general rule? Back A, by a wide margin. Judge a cohort curve by whether it flattens, not by where it starts — the second derivative carries the value, the first point carries the noise.
Business & Marketing basic

Front When does the Rule of 40 stop being meaningful? Back Below roughly $50M ARR. Early companies score 100+ on trivial revenue or deeply negative on deliberate investment. It is a maturity test for scaled recurring-revenue businesses, not a startup filter.
Web Performance & UX basic

Front Lighthouse gives the page 95; CrUX field data fails Core Web Vitals. Which is wrong, and why do they disagree? Back Neither is wrong — they measure different worlds. Lighthouse is one synthetic load: your dev machine or a datacenter, one simulated throttle, an empty cache, no interaction (so no INP at all). Field data is real users on real devices over 28 days — including the mid-range Android on hotel wifi that *is* your p75. Use them by role: field data decides whether you have a problem (it's what ranking uses); lab reproduces and debugs it. Optimizing until Lighthouse is green treats the thermometer instead of the fever.
Business & Marketing basic

Front Terminal growth rate in a DCF — the cap, and why? Back Two to three percent, roughly long-run GDP. Anything higher implies the firm eventually becomes the entire economy.
Backend & Data Engineering basic

Front Memory grows monotonically for days until OOM; a restart resets it. What does the *shape* tell you before you profile anything? Back Monotonic growth is accumulation: an unbounded in-process cache, listeners registered and never removed, per-request state parked in a global. A healthy GC'd process saws — allocation up, collection down. The growth *rate* is the second clue: if it tracks request volume, the leak is per-request; if it's steady regardless of load, look at background work.
Backend & Data Engineering basic

Front Postgres does not index foreign keys automatically. Where does that bite? Back The unique side is indexed; the *referencing* column is not. Two bites: every DELETE on the parent scans the child table to check for references — a parent purge becomes a sequence of full scans — and the join you inevitably write on that FK has no index either. Rule of thumb: creating a foreign key and creating an index on it are one act, separated only by the ORM's failure to do the second.
Backend & Data Engineering basic

Front Latency climbs steadily through the day; a restart fixes it until tomorrow. First hypothesis? Back Something accumulates per unit of work and the restart empties it: heap pressure driving longer GC pauses, a connection or file-descriptor leak approaching a limit, an in-process cache grown past its useful size and thrashing. "Restart fixes it" is diagnostic gold — it localizes the problem to process state and rules out data, schema, and traffic in one observation.
Backend & Data Engineering basic

Front "The database was down for 40 minutes but the queue absorbed everything — no incident." What's the trap in that sentence? Back The debt: 40 minutes of arrivals must now be processed *on top of* live traffic. If consumers normally run at 60% capacity, the spare 40% pays down the backlog at 2/3 the rate it accrued — the queue drains in ~an hour if nothing else goes wrong, and meanwhile every message waits behind the backlog, so end-to-end latency is silently hours even though every component is green. Queues convert an availability incident into a latency incident. That's usually a good trade — but only if queue *age*, not just depth, is on a dashboard with an SLO.
Backend & Data Engineering basic

Front The latency ladder at human scale — if an L1 hit is one second, how long are RAM, SSD, a same-DC call, and a transatlantic call? Back L1 = 1 second → RAM ≈ 2 minutes → SSD random read ≈ a day → same-DC round-trip ≈ 6 days → transatlantic round-trip ≈ 2.5 years. The rescaling is what makes the design instincts visceral: "read it from memory instead of the API" is the difference between two minutes and two years — which is why caching works, why chatty protocols die, and why the CPU spends most of its existence waiting for everything below it.
Business & Marketing basic

Front Renewal rate fell sharply with no change in product or support. First hypothesis? Back Look at what was sold 12 months ago, not at what happened this quarter. A renewal cliff usually reflects an acquisition cohort that was mis-sold, discounted into the wrong segment, or never onboarded — the cause is a year upstream of the symptom.
Business & Marketing basic

Front Website conversion rate doubled overnight with no campaign or product change. First hypothesis? Back Measurement, not performance. A deploy changed event firing, a bot filter was added, or a tag now dedupes. Suspect the instrument before the world: step changes in a ratio almost always come from the denominator.
Business & Marketing basic

Front What does a geo holdout measure that attribution cannot? Back Incrementality — sales that would not have happened anyway. Attribution decides which touchpoint to credit for a conversion; a holdout decides whether the conversion existed at all. On retargeting the two routinely disagree by 2–5x.
Business & Marketing basic

Front "Two months free" on an annual plan — what discount is that, exactly? Back 16.7%, not 20%. You collect ten months of revenue for twelve months of service: 10/12 = 0.833. The common mis-statement ("2 out of 10") understates the concession you are making.
Business & Marketing basic

Front Why does an annual-billing business appear to have almost no churn for its first year? Back Churn can only occur at renewal. A cohort shows 100% retention for eleven months and then takes the entire loss in month twelve. Monthly-cohort intuition applied to annual contracts produces wildly optimistic forecasts right up to the cliff.
Web Performance & UX basic

Front The full-screen app-install interstitial — what does it cost on both channels? Back On SEO: Google explicitly demotes pages showing intrusive interstitials on mobile entry — the banner is spending organic ranking. On users: it interrupts the task *before the site has delivered any value*, at the exact moment intent is most fragile — and the app it pitches mostly serves users who already chose the site. The compliant, better-converting shape: the OS-level smart banner or a dismissible inline strip, shown *after* engagement (second visit, post-purchase) when "get the app" is a service to a proven user rather than a toll on a new one.
Web Performance & UX basic

Front The minimum touch target — the number, and why the *spacing* clause matters as much as the size? Back Roughly 44–48px (Apple says 44pt, Google 48dp) — driven by fingertip contact area, not pixels. And targets need breathing room: a big button flush against another big button still mis-taps at the boundary, which is why the guidelines also specify ~8dp between targets. The cost of missing it isn't abstract: a mis-tap on mobile means an unintended navigation, a back-gesture, a reload — each error costs seconds and trust. The usual offenders are icon rows, list-item chevrons, and text links in running prose.
Backend & Data Engineering basic

Front Cache-aside with delete-on-write is "obviously correct". Describe the race that still serves stale data forever. Back Reader misses, reads value v1 from the DB, and stalls. Writer updates to v2 and deletes the cache key. The stalled reader resumes and writes v1 into the now-empty cache — stale data with nothing left to evict it, cached until the next write or TTL. Rare per-request, guaranteed at volume. The honest defenses: a TTL as backstop (bounding staleness rather than preventing it), or versioned/compare-and-set writes. Full invalidation correctness is famously not worth its complexity for most workloads — pick your staleness bound instead.
Business & Marketing basic

Front Conversion rate on the pricing page fell the week after a redesign, but revenue is flat. First hypothesis? Back The page now filters better, not worse: fewer people click through, but those who do are buying more or converting further down. Read the funnel to revenue before reverting.