全部文章
-
Rust FFI 资源所有权：将外部构造函数与实际对应的析构函数配对
在围绕外部资源构建的窄接口 Rust 包装器中，明确分配、释放和回调生命周期。
-
Bulk operations instead of per-row loops: round trips, transactions and COPY
A loop that sends one statement per row pays a round trip, a parse and often a commit per row; a bulk operation sends the set in one statement, one stream or one transaction. PostgreSQL's own guidance is to commit once, use COPY for loads and build indexes after loading; client libraries offer executemany and copy for the same reason.
-
Network throughput and retransmits on Linux: ss -ti, nstat, and iperf3
ss -ti shows one connection's own retransmission and RTT view; nstat reports the same kind of counters host-wide from the kernel's SNMP and extended TCP statistics; iperf3 measures achievable throughput directly when retransmits are low but throughput still disappoints.
-
A first game day: one chaos experiment with a hypothesis, a blast radius and an abort rule
Run a first fault-injection exercise as a scheduled, announced experiment: define the steady state as a measurable output, hypothesise that it holds under one specific fault, limit the blast radius, set an abort condition, inject, and record what the system and the people did; the Principles of Chaos Engineering give the four steps and the Google SRE book describes disaster role-play as a weekly ritual.
-
Reasoning about complexity before optimising
Asymptotic complexity predicts how run time grows with input size; recognising quadratic loops, repeated linear searches and unbounded recursion in code review prevents most performance incidents before profiling is needed.
-
Separating HDR lighting, exposure, tone mapping, and display output
Use explicit stages to preserve bright scene information, set exposure predictably, and generate a signal appropriate for the target display.
-
A wrist step counter and a pocket phone diverge more on days with many short walking bouts and occupied hands
Hypothesis: two consumer step counters worn by one person, one on the wrist and one in a trouser pocket, disagree by a larger fraction on days the diary marks as many short walks or hands occupied (pram, bags, trolley) than on days with a few long walks, because the two devices read different body motions; a daily log with both devices stated is proposed as the test and no measurement is reported.
-
Protocol Buffers 的确定性序列化不等于规范的消息身份
避免将任意 protobuf 字节用作跨运行时和版本的稳定语义指纹。
-
Separating the acting service from the represented user in delegation tests
Test a delegated action with both the executing service identity and the user it represents visible in the oracle. This proposed method avoids treating successful service authentication as sufficient proof of user authority.
-
HTTP keep-alive and connection reuse: pools, idle timeouts and the stale-connection race
HTTP/1.1 keeps a connection open for further requests unless a Connection: close is sent, which removes a TCP and TLS handshake from every request after the first; the client must keep a pool for the life of the process, read every response body, and set its idle timeout below the server's so it does not reuse a connection the server has already closed.
-
APT pinning: keeping one package on a chosen version or source with preferences files
APT pinning lets an operator hold a single package back, pull it from a non-default suite, or prevent a future apt upgrade from touching it — configured per file under /etc/apt/preferences.d rather than by editing one shared file, and inspected with apt-cache policy.
-
AdamW：解耦权重衰减为什么不同于 L2 惩罚
AdamW 将参数收缩与自适应梯度更新分开，因此衰减项不会进入 Adam 的矩估计；但学习率、更新次数和参数分组仍会影响正则化。
-
Reading, granting, and backing up NTFS permissions with icacls and Get-Acl/Set-Acl
icacls reads and grants NTFS permissions and encodes inheritance as (OI)(CI) flags; icacls /save and /restore back up and reapply an entire ACL tree before a change, and Get-Acl/Set-Acl give the same information as PowerShell objects.
-
Improvements measured after targeting the worst-performing cases are partly regression to the mean
Hypothesis: when an engineering effort selects the worst-scoring endpoints, tests, hosts or tenants by a noisy metric and reports their improvement after intervention, a substantial part of the reported gain would have occurred without the intervention, because extreme measurements tend to be followed by less extreme ones. A proposed test with an untouched control group is described.
-
Shader derivatives: preserve valid gradients across branches and UV seams
Diagnose unstable texture footprints by distinguishing continuous coordinates, explicit gradients and derivative operations in divergent fragment-shader control flow.
-
服务级别目标与错误预算
SLO 是可用性或延迟等服务级别指标的目标值。目标值与 100% 之间的差额就是错误预算，用来决定发布可以承担多大风险。
-
会自我验证的首日设置脚本：bootstrap、doctor 与冒烟测试
克隆后的一条命令，要么生成可用环境，要么给出缺少什么以及如何修复的精确清单：只读 doctor 脚本检查每项要求，幂等 bootstrap 仅安装检查报告缺少的部分，冒烟测试决定退出状态；CI 从干净机器运行它，避免脚本在新人之间的空档期失效。
-
Datenqualitätsprüfungen: Aktualität, Menge, Nullwerte und Eindeutigkeit als Mindestsatz
Vier billige Prüfungen fangen die meisten kaputten Ladeläufe: Ist die Quelle frisch genug, kam eine plausible Zeilenzahl, sind Schlüssel und Kennzahlen gefüllt, ist die erklärte Körnung eindeutig? Jede Prüfung als Abfrage formulieren, die fehlerhafte Zeilen liefert, nach dem Laden und vor dem Veröffentlichen ausführen, Warnung und Blockade trennen.
-
置信区间概览：区间说明什么，又不说明什么
95% 置信区间来自一种在重复抽样中有 95% 的区间覆盖真实值的方法；某个具体区间要么包含真实值，要么不包含。其宽度随样本量平方根的增大而缩小，随离散程度增大而扩大。应将它理解为与数据相容的值的范围，与每个估计值一同报告，并在比较时计算差值的区间。
-
安全解压归档：zip 和 tar 中的路径遍历
归档条目可能使用 ../../etc/passwd 这样的名称、绝对路径或符号链接。只有解析每个目标路径并确认它仍位于目标目录内后，才可解压；拒绝指向目录外的链接，并限制总大小和条目数量。
机器可读： JSON