YouTube video by The Majority Report w/ Sam Seder

YouTube video by ViolentFemmesVEVO

Governments must stop viewing the sharing of chemical data as a surrender of sovereignty and instead appreciate it as the ultimate force multiplier for global security

A customer-support SaaS has a harder constraint than evaluating a feature flag quickly: after a customer reports a missing reply, an incorrect queue assignment, or a conversation shown to the wrong agent group, the team needs enough durable evidence to reconstruct which configuration the request actually received. That constraint changes both the hosting choice and the pricing comparison. **TL;DR:** choose self-hosted or managed feature flags by testing the complete evidence path, not by comparing a headline fee. Preserve the evaluated flag key, non-sensitive subject identifier, variant, configuration revision, evaluation reason, and event time alongside the affected support event; define retention and deletion rules; then measure whether the resulting signals let an operator distinguish a rollout failure from ordinary application noise. Self-hosting gives the team more control over data placement and operations, while managed service transfers more of that operational burden. Neither choice fixes incomplete evidence. ## Should a small support SaaS use self-hosted or managed feature flags? Start with the reconstruction question: can an investigator determine the configuration used for one decision without guessing from the flag's current state? A later snapshot is weak evidence because configuration can change between evaluation and investigation. The useful record joins a business event, such as a conversation assignment, to the flag decision that influenced it. This is the first trade-off. I would require six fields before debating hosting: a stable flag key, a pseudonymous subject key, the returned variant, an immutable configuration revision, the evaluation reason, and a timestamp. The business event needs its own correlation identifier. Sensitive customer data does not belong in a convenient free-form context field; OWASP's logging guidance explicitly calls out data that should usually be excluded or treated carefully, including access tokens, passwords, and sensitive personal data. Keep the payload narrow. A complete dump of targeting context may feel prudent during design review, but it raises exposure, retention, and search costs while burying the decision an investigator needs. Record identifiers that permit an authorized join to the source of truth, then apply access control, retention, and deletion policies to both sides of that join. Evidence without governance becomes another incident. ## Which signals separate a bad rollout from ordinary noise? A flag evaluation counter is useful only when its dimensions remain bounded and its name describes one unit. Prometheus recommends a base unit and warns that every label set creates a new time series. That makes `flag_key` and a small `result` vocabulary plausible metric labels in a controlled catalog; a customer or conversation identifier is not. Per-customer detail belongs in protected logs or traces, where access and retention can be stricter. A practical split is: * Metrics answer whether error, latency, failed-assignment, or reply-failure rates moved by flag and variant. * Traces connect an evaluation to the request or job that consumed it. * Audit records explain who changed configuration, what revision resulted, and when it became eligible for evaluation. * Domain events retain the outcome that matters to the customer and support agent. The split is deliberate. Metrics stay aggregatable; incident evidence stays specific. Do not attach raw targeting attributes to every counter merely because the instrumentation library accepts labels. Cardinality grows from combinations, and a dimension that looks harmless in a test environment can make production queries slow and expensive once workspaces, regions, queues, flags, and variants multiply. Noise wins otherwise. Here is a deliberately small Python shape for the evidence boundary. It validates a fixed schema and uses HMAC-SHA-256 to pseudonymize the internal customer identifier before emission; the surrounding system still needs key management, access control, retention, and a documented way to perform an authorized join. from dataclasses import asdict, dataclass from datetime import datetime, timezone import hashlib import hmac @dataclass(frozen=True) class FlagDecision: event_id: str subject_key: str flag_key: str variant: str config_revision: str reason: str evaluated_at: str def record_decision(*, event_id: str, customer_id: str, flag_key: str, variant: str, config_revision: str, reason: str, pseudonym_key: bytes) -> dict[str, str]: subject_key = hmac.new( pseudonym_key, customer_id.encode("utf-8"), hashlib.sha256 ).hexdigest() decision = FlagDecision( event_id=event_id, subject_key=subject_key, flag_key=flag_key, variant=variant, config_revision=config_revision, reason=reason, evaluated_at=datetime.now(timezone.utc).isoformat(), ) return asdict(decision) This record should be emitted at the point where the application consumes the decision, rather than inferred later from a control-plane change log. Otherwise a cache, stale process, network partition, or asynchronous job can create a gap between intended configuration and observed behavior. ## The hosting choice follows the failure model Self-hosted and managed systems move responsibility; they do not remove it. The relevant comparison is therefore an ownership map with explicit failure tests. A common shortlist frames the search as Flagsmith self-hosted versus Unleash open source versus GrowthBook versus LaunchDarkly pricing. Those names define candidates, not an answer: editions, contracts, and service boundaries can change, while the architectural limitation remains stable. Verify each candidate's current primary documentation and contract against the same evidence test. This article does not rank them because the operational inputs, not a generic product label, determine the result. Decision area | Self-hosted responsibility | Managed-service responsibility | Evidence to request ---|---|---|--- Control-plane availability | Your team operates upgrades, capacity, backups, and recovery | Provider operates the service; your team still designs client behavior | Published limits, recovery objectives, and a tested outage mode Data placement | Your deployment and storage design determine placement | Contract and service architecture constrain placement | Data-flow diagram, retention controls, and export behavior Configuration history | Your storage, backup, and audit design preserve revisions | Service defines audit and history boundaries | Immutable revision identifiers and a documented export Evaluation continuity | Your runtime and cache policy define stale or unavailable behavior | SDK and service contract define it, but application policy remains yours | Tests for cold start, stale cache, timeout, and invalid configuration Operational load | On-call owns the full stack | On-call owns integration and dependency failure | Upgrade drill, restore drill, and incident escalation path The sharp edge is durability. If reconstruction depends on an audit stream that is retained for less time than customer support takes to escalate a report, the system has already discarded the answer. If it depends on querying a mutable current configuration, there was never durable evidence in the first place. Export alone is not proof either: test ordering, duplication, late arrival, schema evolution, and restoration into a clean environment. Self-hosting is unsuitable when nobody owns upgrades, restore drills, and security response; managed service is unsuitable when its verified data-placement, retention, export, or unavailable-control-plane behavior conflicts with the system's requirements. These are limitations, not footnotes, and a cheaper quote cannot cancel them. Cost belongs in this table, but as an operational model rather than a teaser price. For self-hosting, count compute, replicated storage, backups, upgrades, security response, observability, and on-call time. For managed service, count the contracted usage dimensions, retention, data export, support, and the engineering work required to tolerate dependency failure. Use your measured evaluation volume, environment count, retention period, and recovery target. A generic "small SaaS" label is too vague to produce a defensible answer. ## Test the evidence path before choosing Run the same acceptance exercise against every candidate. Configure 2 variants, generate 20 known conversation-assignment events, change the configuration revision once, interrupt control-plane access, restart an evaluator with an empty cache, and deliver 5 duplicate events out of order. Then ask an operator who did not design the test to reconstruct the decision for 1 event. These counts are test inputs, not performance claims; their purpose is to make missing, duplicated, and stale evidence visible in a repeatable fixture. Score the result on signal quality: was the exact revision identifiable, were timestamps comparable, did duplicate delivery remain recognizable, and could the operator distinguish an evaluation fallback from a normal variant? Also score noise: series count, log volume, fields with no investigative use, and alerts that fired without an actionable symptom. Be strict here. A dashboard that shows a correlated spike is a lead, not a reconstruction. The evidence chain should connect configuration revision, evaluated result, consuming event, and customer-visible outcome, while the audit trail identifies the configuration change separately. Deployment deserves the same skepticism. Put schema validation in continuous integration, exercise fallback semantics in integration tests, canary new instrumentation, and alert on missing evidence rather than raw evaluation volume. An evidence pipeline can report healthy throughput while silently dropping the revision field that makes its records useful. ## Roll out the decision in reversible steps Begin with one high-impact support workflow and a short, documented evidence schema. Establish baseline event volume and metric cardinality, verify access controls, and perform a timed reconstruction exercise. Then extend coverage by risk, not by flag count. Only after the evidence path passes should the team migrate more flags or commit to a hosting model. Keep configuration exports and application behavior tests portable, document the unavailable-control-plane policy, and rehearse restoration. The right choice is the one whose failure modes the team can detect, whose evidence it can retain responsibly, and whose operational obligations it will actually staff. ## Sources * https://prometheus.io/docs/practices/naming/ * https://cheatsheetseries.owasp.org/cheatsheets/Logging_Cheat_Sheet.html

Magent Forensics, the owner of the GrayKey phone unlocking tool, says it can bypass an iPhone rebooting feature that was locking cops out.

Great writing depends on great editing—and great editing is a skill that can be learned. In this highly interactive 5-week course, participants will develop practical editing techniques that…

Many police departments and federal agencies use body language analysis as a forensics technique, claiming that these tools can help assess people’s intentions or truthfulness.

The DistroWatch news feed is brought to you by TUXEDO COMPUTERS. Parrot is a Debian-based, security-oriented distribution featuring a collection of utilities designed for penetration testing, computer forensics, reverse engineering, hacking, privacy, anonymity and cryptography. The project's latest release, version 7.4, includes several changes, including a new onboarding tool: "A new tool has been added to our collection -....

The package originated in Entebbe, Uganda, bound for Cairo, containing 10 kilograms of thumb-sized pieces that could have been either bone or elephant ivory.

They're perfectly to scale, and all based on real cases.

Bruce66423 shares a report from The Telegraph: British police forces, including specialist units within the Metropolitan Police, have used Oxygen Forensics software to break into the phones of suspect...